@qvac/tts-ggml 0.13.0 → 0.14.0

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

@@ -34,6 +34,7 @@ const {
 const { resolveRefWavPath } = require('../utils/runChatterboxTTS')
 const { TTS_TEST_THREADS } = require('../utils/testThreads')
 const { GPU_ONLY_USE_GPU } = require('../utils/gpuOnly')
+const { assertDesktopGpuBackend } = require('../utils/desktopGpuBackend')
 
 const platform = os.platform()
 const isMobile = platform === 'ios' || platform === 'android'
@@ -42,11 +43,6 @@ const isMobile = platform === 'ios' || platform === 'android'
 // fallback from a failure to a warning (e.g. a Linux host with no Vulkan SDK).
 const NO_GPU = proc.env && proc.env.NO_GPU === 'true'
 const RELAX = proc.env && proc.env.QVAC_TTS_GPU_SMOKE_RELAX === '1'
-const GPU_BACKEND_IDS = { metal: 1, cuda: 2, vulkan: 3, opencl: 4 }
-// CI rows whose prebuild bundles more than one usable GPU backend (the linux
-// runners carry CUDA and Vulkan) pin the engine cascade through
-// TTS_CPP_GPU_BACKEND; the enhancer assertion expects whatever the row pinned
-// and falls back to the platform's cascade default when unset.
 const PINNED_GPU_BACKEND = (proc.env && proc.env.TTS_CPP_GPU_BACKEND) || ''
 
 const ENHANCED_RATE = 48000
@@ -92,10 +88,6 @@ function expectsGpu() {
   )
 }
 
-function expectedDesktopEnhancerBackendId() {
-  return GPU_BACKEND_IDS[PINNED_GPU_BACKEND] || GPU_BACKEND_IDS.vulkan
-}
-
 // Strict check that the LavaSR *enhancer* network engaged the GPU. Reads the
 // enhancer-specific backend surfaced by ChatterboxModel/SupertonicModel
 // runtimeStats (enhancerBackendDevice: -1 none / 0 CPU / 1 GPU;
@@ -134,12 +126,7 @@ function assertEnhancerGpuBackend(t, stats) {
   if (platform === 'darwin' || platform === 'ios') {
     t.is(id, 1, `enhancer/${platform}: expected Metal backendId=1, got ${backendIdToName(id)}`)
   } else if (platform === 'linux' || platform === 'win32') {
-    const expectedId = expectedDesktopEnhancerBackendId()
-    t.is(
-      id,
-      expectedId,
-      `enhancer/${platform}: expected ${backendIdToName(expectedId)} backendId=${expectedId}, got ${backendIdToName(id)}`
-    )
+    assertDesktopGpuBackend(t, `enhancer/${platform}`, dev, id, PINNED_GPU_BACKEND)
   } else if (platform === 'android') {
     t.ok(
       id === 3 || id === 4,

package/test/integration/moss-sfx.test.js

@@ -1,10 +1,12 @@
 'use strict'
 
 const path = require('bare-path')
+const fs = require('bare-fs')
 const proc = require('bare-process')
 const test = require('brittle')
 
 const TTSGgml = require('@qvac/tts-ggml')
+const { assertMossBackend } = require('../utils/mossBackend')
 
 const SFX_SAMPLE_RATE = 48000
 const TENTHS_PER_SECOND = 10
@@ -14,13 +16,35 @@ const MODEL_DIR_ENV = 'QVAC_TEST_MOSS_SFX_MODEL_DIR'
 const QUICK_STEPS = 2
 const SHORT_SECONDS = 1.5
 const LONGER_SECONDS = 2.5
-const CPU_DEVICE = 0
-const GPU_DEVICE = 1
+const PINNED_GPU_BACKEND = (proc.env && proc.env.TTS_CPP_GPU_BACKEND) || ''
 
 const modelDir = (proc.env && proc.env[MODEL_DIR_ENV]) || ''
 const skipWithoutModels = modelDir === ''
 const noGpu = proc.env && proc.env.NO_GPU === 'true'
 
+if (proc.env && proc.env.QVAC_TEST_MOSS_FULL_GPU === 'true') {
+  if (noGpu || !['cuda', 'vulkan', 'metal', 'opencl'].includes(PINNED_GPU_BACKEND.toLowerCase())) {
+    throw new Error(
+      'MOSS-SFX full GPU validation requires an explicit GPU backend and NO_GPU=false'
+    )
+  }
+  if (
+    !modelDir ||
+    !fs.existsSync(modelDir) ||
+    !fs
+      .readdirSync(modelDir)
+      .some(
+        (name) =>
+          /^moss-sfx(-[a-z0-9_]+)*\.gguf$/i.test(name) &&
+          fs.statSync(path.join(modelDir, name)).isFile()
+      )
+  ) {
+    throw new Error(
+      'MOSS-SFX full GPU validation requires QVAC_TEST_MOSS_SFX_MODEL_DIR with a GGUF'
+    )
+  }
+}
+
 function createSfxModel(useGPU = false) {
   return new TTSGgml({
     engine: TTSGgml.ENGINE_MOSS_SFX,
@@ -70,7 +94,7 @@ test(
     try {
       const result = await generate(model, { seconds: SHORT_SECONDS })
       assertClip(t, 'cpu clip', result, SHORT_SECONDS)
-      t.is(result.stats.backendDevice, CPU_DEVICE, 'the clip ran on the CPU')
+      assertMossBackend(t, 'cpu clip', result.stats, false, PINNED_GPU_BACKEND)
     } finally {
       await model.unload()
     }
@@ -88,7 +112,8 @@ test(
       const longer = await generate(model, { seconds: LONGER_SECONDS, negativePrompt: 'music' })
       assertClip(t, 'short clip', short, SHORT_SECONDS)
       assertClip(t, 'longer clip', longer, LONGER_SECONDS)
-      t.is(short.stats.backendDevice, GPU_DEVICE, 'the clips ran on the GPU')
+      assertMossBackend(t, 'short clip', short.stats, true, PINNED_GPU_BACKEND)
+      assertMossBackend(t, 'longer clip', longer.stats, true, PINNED_GPU_BACKEND)
     } finally {
       await model.unload()
     }

package/test/integration/moss.test.js

@@ -290,7 +290,7 @@ test(
         { dialogueReferences: references, seed: 7, files: { modelDir: path.resolve(modelDir) } },
         async (model) => {
           t.ok(
-            DIALOGUE_BACKBONE_RE.test(path.basename(model._mossBackbonePath)),
+            DIALOGUE_BACKBONE_RE.test(path.basename(model._setup.paths.mossBackbonePath)),
             'dialogue picks the TTSD backbone'
           )
           const started = Date.now()

package/test/utils/desktopGpuBackend.js

@@ -0,0 +1,21 @@
+'use strict'
+
+const GPU_DEVICE = 1
+const GPU_BACKEND_IDS = { metal: 1, cuda: 2, vulkan: 3, opencl: 4 }
+const AUTOMATIC_BACKEND_IDS = [GPU_BACKEND_IDS.cuda, GPU_BACKEND_IDS.vulkan]
+
+function expectedBackendIds(pinnedBackend) {
+  if (!pinnedBackend) return AUTOMATIC_BACKEND_IDS
+  const name = pinnedBackend.toLowerCase()
+  return Object.prototype.hasOwnProperty.call(GPU_BACKEND_IDS, name) ? [GPU_BACKEND_IDS[name]] : []
+}
+
+function assertDesktopGpuBackend(t, label, device, backendId, pinnedBackend = '') {
+  t.is(device, GPU_DEVICE, `${label}: GPU device required`)
+  t.ok(
+    expectedBackendIds(pinnedBackend).includes(backendId),
+    `${label}: expected ${pinnedBackend || 'CUDA or Vulkan'} backend, got backendId=${backendId}`
+  )
+}
+
+module.exports = { assertDesktopGpuBackend }

package/test/utils/runTTS.js

@@ -58,7 +58,7 @@ async function runTTSWithSplit(model, params, expectation = {}, options = {}) {
     const splitLanguage =
       typeof params.splitLanguage === 'string' && params.splitLanguage.length > 0
         ? params.splitLanguage
-        : model?._config?.language || 'en'
+        : model?._setup?.config?.language || 'en'
     const splitOpts = {
       language: splitLanguage,
       mergeToMaxScalars: false,