@qvac/inference 0.19.1 → 0.20.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (137) hide show
  1. package/NOTICE +43 -11
  2. package/dist/api/audio-gen.d.ts +42 -1
  3. package/dist/api/audio-gen.js +137 -51
  4. package/dist/api/index.d.ts +2 -1
  5. package/dist/api/index.js +2 -1
  6. package/dist/api/text-to-speech.d.ts +31 -4
  7. package/dist/api/text-to-speech.js +117 -21
  8. package/dist/api/vector-index.d.ts +142 -0
  9. package/dist/api/vector-index.js +168 -0
  10. package/dist/api/world.d.ts +2 -2
  11. package/dist/errors/index.d.ts +12 -0
  12. package/dist/errors/index.js +21 -0
  13. package/dist/handlers/vector-index.d.ts +2 -0
  14. package/dist/handlers/vector-index.js +40 -0
  15. package/dist/index.d.ts +1 -1
  16. package/dist/index.js +1 -1
  17. package/dist/models/registry/models.d.ts +351 -12
  18. package/dist/models/registry/models.js +2271 -1957
  19. package/dist/models/registry/resource-profiles.js +135 -0
  20. package/dist/plugins/builtin/audiogen-ggml/ops/audio-edit-stream.d.ts +2 -0
  21. package/dist/plugins/builtin/audiogen-ggml/ops/audio-edit-stream.js +65 -0
  22. package/dist/plugins/builtin/audiogen-ggml/ops/audio-gen-input.d.ts +9 -1
  23. package/dist/plugins/builtin/audiogen-ggml/ops/audio-gen-input.js +24 -7
  24. package/dist/plugins/builtin/audiogen-ggml/ops/audio-gen-run.d.ts +46 -0
  25. package/dist/plugins/builtin/audiogen-ggml/ops/audio-gen-run.js +218 -0
  26. package/dist/plugins/builtin/audiogen-ggml/ops/audio-gen-stream.d.ts +1 -1
  27. package/dist/plugins/builtin/audiogen-ggml/ops/audio-gen-stream.js +56 -225
  28. package/dist/plugins/builtin/audiogen-ggml/ops/audio-understand.d.ts +8 -0
  29. package/dist/plugins/builtin/audiogen-ggml/ops/audio-understand.js +28 -0
  30. package/dist/plugins/builtin/audiogen-ggml/plugin.d.ts +248 -0
  31. package/dist/plugins/builtin/audiogen-ggml/plugin.js +17 -1
  32. package/dist/plugins/builtin/bci-whispercpp-transcription/plugin.d.ts +5 -2
  33. package/dist/plugins/builtin/bci-whispercpp-transcription/plugin.js +1 -2
  34. package/dist/plugins/builtin/ggml-ocr/plugin.d.ts +1 -1
  35. package/dist/plugins/builtin/ggml-ocr/plugin.js +1 -2
  36. package/dist/plugins/builtin/llamacpp-completion/ops/batch-completion-stream.js +6 -4
  37. package/dist/plugins/builtin/llamacpp-completion/ops/cache-logger.d.ts +1 -2
  38. package/dist/plugins/builtin/llamacpp-completion/ops/cache-logger.js +0 -4
  39. package/dist/plugins/builtin/llamacpp-completion/ops/completion-stats.d.ts +1 -0
  40. package/dist/plugins/builtin/llamacpp-completion/ops/completion-stats.js +3 -0
  41. package/dist/plugins/builtin/llamacpp-completion/ops/completion-stream.d.ts +9 -0
  42. package/dist/plugins/builtin/llamacpp-completion/ops/completion-stream.js +96 -89
  43. package/dist/plugins/builtin/llamacpp-completion/ops/context-overflow.d.ts +0 -1
  44. package/dist/plugins/builtin/llamacpp-completion/ops/context-overflow.js +0 -31
  45. package/dist/plugins/builtin/llamacpp-completion/ops/kv-cache-session.d.ts +40 -24
  46. package/dist/plugins/builtin/llamacpp-completion/ops/kv-cache-session.js +187 -123
  47. package/dist/plugins/builtin/llamacpp-completion/ops/kv-cache-state.d.ts +11 -4
  48. package/dist/plugins/builtin/llamacpp-completion/ops/kv-cache-state.js +18 -16
  49. package/dist/plugins/builtin/llamacpp-completion/plugin.d.ts +8 -4
  50. package/dist/plugins/builtin/llamacpp-completion/plugin.js +1 -2
  51. package/dist/plugins/builtin/llamacpp-embedding/plugin.d.ts +1 -2
  52. package/dist/plugins/builtin/llamacpp-embedding/plugin.js +2 -38
  53. package/dist/plugins/builtin/llamacpp-embedding/transform.d.ts +9 -0
  54. package/dist/plugins/builtin/llamacpp-embedding/transform.js +42 -0
  55. package/dist/plugins/builtin/nmtcpp-translation/plugin.d.ts +1 -1
  56. package/dist/plugins/builtin/nmtcpp-translation/plugin.js +1 -2
  57. package/dist/plugins/builtin/parakeet-transcription/plugin.d.ts +1 -1
  58. package/dist/plugins/builtin/parakeet-transcription/plugin.js +1 -2
  59. package/dist/plugins/builtin/sdcpp-generation/ops/video.d.ts +1 -0
  60. package/dist/plugins/builtin/sdcpp-generation/ops/video.js +8 -2
  61. package/dist/plugins/builtin/sdcpp-generation/plugin.d.ts +9 -4
  62. package/dist/plugins/builtin/sdcpp-generation/plugin.js +51 -19
  63. package/dist/plugins/builtin/tts-ggml/ops/cancel-binding.d.ts +29 -0
  64. package/dist/plugins/builtin/tts-ggml/ops/cancel-binding.js +66 -0
  65. package/dist/plugins/builtin/tts-ggml/ops/parler-options.js +10 -0
  66. package/dist/plugins/builtin/tts-ggml/ops/text-to-speech-stream.d.ts +3 -5
  67. package/dist/plugins/builtin/tts-ggml/ops/text-to-speech-stream.js +143 -28
  68. package/dist/plugins/builtin/tts-ggml/ops/text-to-speech.d.ts +5 -2
  69. package/dist/plugins/builtin/tts-ggml/ops/text-to-speech.js +182 -66
  70. package/dist/plugins/builtin/tts-ggml/plugin.d.ts +267 -22
  71. package/dist/plugins/builtin/tts-ggml/plugin.js +94 -35
  72. package/dist/plugins/builtin/whispercpp-transcription/plugin.d.ts +1 -1
  73. package/dist/plugins/builtin/whispercpp-transcription/plugin.js +1 -2
  74. package/dist/plugins/index.d.ts +1 -1
  75. package/dist/plugins/index.js +1 -1
  76. package/dist/plugins/ops/load-model.js +22 -2
  77. package/dist/plugins/ops/translate.js +58 -12
  78. package/dist/plugins/ops/translatepsy-utils.d.ts +17 -0
  79. package/dist/plugins/ops/translatepsy-utils.js +32 -0
  80. package/dist/plugins/registry.d.ts +15 -0
  81. package/dist/plugins/registry.js +133 -15
  82. package/dist/plugins/turbovec-backend.d.ts +17 -0
  83. package/dist/plugins/turbovec-backend.js +99 -0
  84. package/dist/registry.js +4 -0
  85. package/dist/resources/model-fit/native-probe/advisory-fit.d.ts +46 -0
  86. package/dist/resources/model-fit/native-probe/advisory-fit.js +214 -0
  87. package/dist/resources/model-fit/native-probe/create-llama-fit-request.d.ts +26 -0
  88. package/dist/resources/model-fit/native-probe/create-llama-fit-request.js +171 -0
  89. package/dist/resources/model-fit/native-probe/run-in-process-fit.d.ts +42 -0
  90. package/dist/resources/model-fit/native-probe/run-in-process-fit.js +192 -0
  91. package/dist/resources/model-fit/native-probe/run-isolated-fit.d.ts +70 -0
  92. package/dist/resources/model-fit/native-probe/run-isolated-fit.js +372 -0
  93. package/dist/runtime/env.d.ts +2 -0
  94. package/dist/runtime/env.js +9 -2
  95. package/dist/runtime/lifecycle.js +2 -0
  96. package/dist/runtime/model-registry.d.ts +3 -1
  97. package/dist/runtime/model-registry.js +2 -1
  98. package/dist/runtime/request-context.js +17 -0
  99. package/dist/runtime/vector-index-backend.d.ts +25 -0
  100. package/dist/runtime/vector-index-backend.js +1 -0
  101. package/dist/runtime/vector-index-registry.d.ts +76 -0
  102. package/dist/runtime/vector-index-registry.js +161 -0
  103. package/dist/schemas/assess-model-fit.d.ts +44 -0
  104. package/dist/schemas/assess-model-fit.js +40 -0
  105. package/dist/schemas/audio-gen.d.ts +479 -4
  106. package/dist/schemas/audio-gen.js +376 -17
  107. package/dist/schemas/batch-completion-stream.d.ts +7 -0
  108. package/dist/schemas/batch-completion-stream.js +5 -3
  109. package/dist/schemas/common.d.ts +749 -15
  110. package/dist/schemas/common.js +8 -1
  111. package/dist/schemas/completion-event.d.ts +11 -1
  112. package/dist/schemas/completion-event.js +3 -0
  113. package/dist/schemas/completion-stream.d.ts +20 -0
  114. package/dist/schemas/completion-stream.js +41 -4
  115. package/dist/schemas/config.d.ts +48 -30
  116. package/dist/schemas/errors.d.ts +4 -0
  117. package/dist/schemas/errors.js +22 -0
  118. package/dist/schemas/index.d.ts +1 -0
  119. package/dist/schemas/index.js +1 -0
  120. package/dist/schemas/llamacpp-config.d.ts +4 -8
  121. package/dist/schemas/llamacpp-config.js +4 -4
  122. package/dist/schemas/load-model.d.ts +2503 -139
  123. package/dist/schemas/model-config-schemas.d.ts +228 -5
  124. package/dist/schemas/plugin.d.ts +10 -0
  125. package/dist/schemas/sdcpp-config.d.ts +221 -4
  126. package/dist/schemas/sdcpp-config.js +128 -22
  127. package/dist/schemas/text-to-speech.d.ts +780 -19
  128. package/dist/schemas/text-to-speech.js +298 -44
  129. package/dist/schemas/transcription-config.js +7 -4
  130. package/dist/schemas/vector-index.d.ts +194 -0
  131. package/dist/schemas/vector-index.js +185 -0
  132. package/dist/utils/addon-responses.d.ts +29 -3
  133. package/dist/utils/aggregate-events.d.ts +3 -1
  134. package/dist/utils/aggregate-events.js +7 -1
  135. package/dist/utils/tts-stats.d.ts +15 -11
  136. package/dist/utils/tts-stats.js +33 -11
  137. package/package.json +28 -25
@@ -1,4 +1,3 @@
1
- import ttsAddonLogging from '@qvac/tts-ggml/addonLogging';
2
1
  import TTSGgml from '@qvac/tts-ggml';
3
2
  import { definePlugin, defineHandler, defineDuplexHandler, ttsRequestSchema, ttsResponseSchema, textToSpeechStreamRequestSchema, textToSpeechStreamResponseSchema, ModelType, ttsConfigSchema, ADDON_TTS, LEGACY_TTS_ONNX_MODEL_CONFIG_FIELDS } from '../../../schemas/index.js';
4
3
  import { createStreamLogger, registerAddonLogger } from '../../../logging/index.js';
@@ -6,6 +5,19 @@ import { TtsArtifactsRequiredError, LegacyTtsModelDeprecatedError } from '../../
6
5
  import { textToSpeech } from './ops/text-to-speech.js';
7
6
  import { textToSpeechStream } from './ops/text-to-speech-stream.js';
8
7
  import { attachModelExecutionMs } from '../../../profiling/model-execution.js';
8
+ // Everything a TTS frame carries beyond its PCM, forwarded verbatim from the
9
+ // op yield onto the wire frame. Both handlers use it so the two response
10
+ // shapes cannot drift apart.
11
+ function ttsFrameMetadata(value) {
12
+ return {
13
+ ...(value.sampleRate !== undefined ? { sampleRate: value.sampleRate } : {}),
14
+ ...(value.chunkIndex !== undefined ? { chunkIndex: value.chunkIndex } : {}),
15
+ ...(typeof value.sentenceChunk === 'string' && value.sentenceChunk.length > 0
16
+ ? { sentenceChunk: value.sentenceChunk }
17
+ : {}),
18
+ ...(value.isLast !== undefined ? { isLast: value.isLast } : {})
19
+ };
20
+ }
9
21
  function rejectLegacyOnnxFields(cfg) {
10
22
  const record = cfg;
11
23
  const legacyFields = LEGACY_TTS_ONNX_MODEL_CONFIG_FIELDS.filter((name) => record[name] !== undefined);
@@ -63,17 +75,34 @@ async function resolveSupertonicConfig(config, ctx) {
63
75
  const lavasrArtifacts = await resolveLavasrArtifacts(lavasrEnhancerModelSrc, lavasrDenoiserModelSrc, ctx);
64
76
  return { config: runtime, artifacts: lavasrArtifacts };
65
77
  }
66
- function resolveParlerConfig(config) {
67
- return { config, artifacts: {} };
68
- }
69
- // CosyVoice3's multi-file layout (flow/HiFT GGUFs, voice.gguf, vocab.json,
70
- // merges.txt) rides the primary model's registry companion set, so only the
71
- // optional LavaSR post-processing sources need resolution here.
72
- async function resolveCosyvoice3Config(config, ctx) {
78
+ async function resolveParlerConfig(config, ctx) {
73
79
  const { lavasrEnhancerModelSrc, lavasrDenoiserModelSrc, ...runtime } = config;
74
80
  const lavasrArtifacts = await resolveLavasrArtifacts(lavasrEnhancerModelSrc, lavasrDenoiserModelSrc, ctx);
75
81
  return { config: runtime, artifacts: lavasrArtifacts };
76
82
  }
83
+ // CosyVoice3's multi-file layout (flow/HiFT GGUFs, voice.gguf, vocab.json,
84
+ // merges.txt) rides the primary model's registry companion set. The voice
85
+ // cloning add-ons are deliberately not in that set — they are a ~300 MB opt-in
86
+ // — so they are resolved here alongside the optional LavaSR sources.
87
+ async function resolveCosyvoice3Config(config, ctx) {
88
+ const { lavasrEnhancerModelSrc, lavasrDenoiserModelSrc, referenceAudioSrc, cosyvoice3S3tokModelSrc, cosyvoice3CampplusModelSrc, ...runtime } = config;
89
+ const resolve = ctx.resolveModelPath;
90
+ const [lavasrArtifacts, referenceAudioPath, cosyvoiceS3tokPath, cosyvoiceCampplusPath] = await Promise.all([
91
+ resolveLavasrArtifacts(lavasrEnhancerModelSrc, lavasrDenoiserModelSrc, ctx),
92
+ referenceAudioSrc ? resolve(referenceAudioSrc) : Promise.resolve(undefined),
93
+ cosyvoice3S3tokModelSrc ? resolve(cosyvoice3S3tokModelSrc) : Promise.resolve(undefined),
94
+ cosyvoice3CampplusModelSrc ? resolve(cosyvoice3CampplusModelSrc) : Promise.resolve(undefined)
95
+ ]);
96
+ return {
97
+ config: runtime,
98
+ artifacts: {
99
+ ...lavasrArtifacts,
100
+ ...(referenceAudioPath ? { referenceAudioPath } : {}),
101
+ ...(cosyvoiceS3tokPath ? { cosyvoiceS3tokPath } : {}),
102
+ ...(cosyvoiceCampplusPath ? { cosyvoiceCampplusPath } : {})
103
+ }
104
+ };
105
+ }
77
106
  async function resolveAudio8Config(config, ctx) {
78
107
  const { audio8CodecDecoderModelSrc, audio8CodecEncoderModelSrc, referenceAudioSrc, ...runtime } = config;
79
108
  if (!audio8CodecDecoderModelSrc) {
@@ -94,6 +123,16 @@ async function resolveAudio8Config(config, ctx) {
94
123
  }
95
124
  };
96
125
  }
126
+ // The generic ggml backend-loading knobs, forwarded on `config` the same way
127
+ // Supertonic's vulkanCacheDir already is. Parler and Audio8 configs carry no
128
+ // `openclCacheDir` (their native builders do not read it), so it is simply
129
+ // absent for them.
130
+ function ggmlBackendConfig(config) {
131
+ return {
132
+ ...(config.backendsDir !== undefined ? { backendsDir: config.backendsDir } : {}),
133
+ ...(config.openclCacheDir !== undefined ? { openclCacheDir: config.openclCacheDir } : {})
134
+ };
135
+ }
97
136
  // Build the optional LavaSR `files` entries from resolved artifacts. Supplying
98
137
  // a path is what enables the stage in @qvac/tts-ggml — there is no on/off flag.
99
138
  function lavasrFiles(artifacts) {
@@ -132,12 +171,20 @@ function createChatterboxModel(modelId, config, params, artifacts) {
132
171
  : {}),
133
172
  ...(config.cfmSteps !== undefined ? { cfmSteps: config.cfmSteps } : {}),
134
173
  ...(config.cfgRate !== undefined ? { cfgRate: config.cfgRate } : {}),
174
+ ...(config.nCtx !== undefined ? { nCtx: config.nCtx } : {}),
175
+ ...(config.kvCacheType !== undefined ? { kvCacheType: config.kvCacheType } : {}),
176
+ // Chatterbox has no pace channel; `speed` is its only rate control.
177
+ ...(config.ttsSpeed !== undefined ? { speed: config.ttsSpeed } : {}),
135
178
  ...(config.threads !== undefined ? { threads: config.threads } : {}),
136
179
  ...(config.nGpuLayers !== undefined ? { nGpuLayers: config.nGpuLayers } : {}),
137
180
  ...(config.seed !== undefined ? { seed: config.seed } : {}),
138
181
  config: {
139
182
  language: config.language ?? 'en',
140
- ...(config.useGPU !== undefined ? { useGPU: config.useGPU } : {})
183
+ ...(config.useGPU !== undefined ? { useGPU: config.useGPU } : {}),
184
+ ...(config.outputSampleRate !== undefined
185
+ ? { outputSampleRate: config.outputSampleRate }
186
+ : {}),
187
+ ...ggmlBackendConfig(config)
141
188
  },
142
189
  logger,
143
190
  opts: { stats: true },
@@ -157,16 +204,26 @@ function createSupertonicModel(modelId, config, params, artifacts) {
157
204
  files: { supertonicModel, ...lavasrFiles(artifacts) },
158
205
  voice: config.voice ?? 'F1',
159
206
  ...(config.ttsSpeed !== undefined ? { speed: config.ttsSpeed } : {}),
207
+ // Supertonic conditions pace when the engine is built, so it arrives as a
208
+ // constructor option rather than a per-request field.
209
+ ...(config.pace !== undefined ? { pace: config.pace } : {}),
160
210
  ...(config.ttsNumInferenceSteps !== undefined
161
211
  ? { numInferenceSteps: config.ttsNumInferenceSteps }
162
212
  : {}),
213
+ ...(config.threads !== undefined ? { threads: config.threads } : {}),
214
+ ...(config.nGpuLayers !== undefined ? { nGpuLayers: config.nGpuLayers } : {}),
215
+ ...(config.seed !== undefined ? { seed: config.seed } : {}),
163
216
  config: {
164
217
  language: config.language ?? 'en',
165
- useGPU: config.useGPU ?? false,
218
+ // Supertonic defaults to CPU, but the addon rejects a stated `useGPU`
219
+ // that contradicts `nGpuLayers`. Deriving the default from the layer
220
+ // count keeps `nGpuLayers: 99` on its own usable instead of throwing.
221
+ useGPU: config.useGPU ?? (config.nGpuLayers !== undefined && config.nGpuLayers !== 0),
166
222
  ...(config.outputSampleRate !== undefined
167
223
  ? { outputSampleRate: config.outputSampleRate }
168
224
  : {}),
169
- ...(config.vulkanCacheDir !== undefined ? { vulkanCacheDir: config.vulkanCacheDir } : {})
225
+ ...(config.vulkanCacheDir !== undefined ? { vulkanCacheDir: config.vulkanCacheDir } : {}),
226
+ ...ggmlBackendConfig(config)
170
227
  },
171
228
  logger,
172
229
  opts: { stats: true },
@@ -175,7 +232,7 @@ function createSupertonicModel(modelId, config, params, artifacts) {
175
232
  registerAddonLogger(modelId, ModelType.ttsGgml, logger);
176
233
  return { model };
177
234
  }
178
- function createParlerModel(modelId, config, params) {
235
+ function createParlerModel(modelId, config, params, artifacts) {
179
236
  const parlerModel = params.modelPath;
180
237
  if (!parlerModel) {
181
238
  throw new TtsArtifactsRequiredError();
@@ -183,7 +240,7 @@ function createParlerModel(modelId, config, params) {
183
240
  const logger = createStreamLogger(modelId, ModelType.ttsGgml);
184
241
  const model = new TTSGgml({
185
242
  engine: TTSGgml.ENGINE_PARLER,
186
- files: { parlerModel },
243
+ files: { parlerModel, ...lavasrFiles(artifacts) },
187
244
  ...(config.description !== undefined ? { description: config.description } : {}),
188
245
  ...(config.voiceDescription !== undefined ? { voiceDescription: config.voiceDescription } : {}),
189
246
  ...(config.voice !== undefined ? { voice: config.voice } : {}),
@@ -213,7 +270,8 @@ function createParlerModel(modelId, config, params) {
213
270
  ...(config.useGPU !== undefined ? { useGPU: config.useGPU } : {}),
214
271
  ...(config.outputSampleRate !== undefined
215
272
  ? { outputSampleRate: config.outputSampleRate }
216
- : {})
273
+ : {}),
274
+ ...ggmlBackendConfig(config)
217
275
  },
218
276
  logger,
219
277
  opts: { stats: true },
@@ -238,6 +296,9 @@ function createCosyvoice3Model(modelId, config, params, artifacts) {
238
296
  if (!cosyvoiceLlmModel) {
239
297
  throw new TtsArtifactsRequiredError();
240
298
  }
299
+ const referenceAudioPath = artifacts['referenceAudioPath'];
300
+ const cosyvoiceS3tokModel = artifacts['cosyvoiceS3tokPath'];
301
+ const cosyvoiceCampplusModel = artifacts['cosyvoiceCampplusPath'];
241
302
  const logger = createStreamLogger(modelId, ModelType.ttsGgml);
242
303
  const model = new TTSGgml({
243
304
  engine: TTSGgml.ENGINE_COSYVOICE3,
@@ -247,8 +308,14 @@ function createCosyvoice3Model(modelId, config, params, artifacts) {
247
308
  files: {
248
309
  cosyvoiceModelDir: dirname(cosyvoiceLlmModel),
249
310
  cosyvoiceLlmModel,
311
+ // Cloning add-ons live outside the companion set, so pass explicit paths
312
+ // rather than relying on the addon's name-prefix discovery in the dir.
313
+ ...(cosyvoiceS3tokModel ? { cosyvoiceS3tokModel } : {}),
314
+ ...(cosyvoiceCampplusModel ? { cosyvoiceCampplusModel } : {}),
250
315
  ...lavasrFiles(artifacts)
251
316
  },
317
+ ...(referenceAudioPath ? { referenceAudio: referenceAudioPath } : {}),
318
+ ...(config.promptText !== undefined ? { promptText: config.promptText } : {}),
252
319
  ...(config.emotion !== undefined ? { emotion: config.emotion } : {}),
253
320
  ...(config.pace !== undefined ? { pace: config.pace } : {}),
254
321
  ...(config.instruct !== undefined ? { instruct: toAddonInstruct(config.instruct) } : {}),
@@ -265,7 +332,8 @@ function createCosyvoice3Model(modelId, config, params, artifacts) {
265
332
  ...(config.useGPU !== undefined ? { useGPU: config.useGPU } : {}),
266
333
  ...(config.outputSampleRate !== undefined
267
334
  ? { outputSampleRate: config.outputSampleRate }
268
- : {})
335
+ : {}),
336
+ ...ggmlBackendConfig(config)
269
337
  },
270
338
  logger,
271
339
  opts: { stats: true },
@@ -304,7 +372,8 @@ function createAudio8Model(modelId, config, params, artifacts) {
304
372
  ...(config.useGPU !== undefined ? { useGPU: config.useGPU } : {}),
305
373
  ...(config.outputSampleRate !== undefined
306
374
  ? { outputSampleRate: config.outputSampleRate }
307
- : {})
375
+ : {}),
376
+ ...ggmlBackendConfig(config)
308
377
  },
309
378
  logger,
310
379
  opts: { stats: true },
@@ -322,7 +391,7 @@ export const ttsPlugin = definePlugin({
322
391
  const { ttsEngine } = cfg;
323
392
  // Same default as the former onnx-tts plugin: omitting `ttsEngine` → Chatterbox.
324
393
  if (ttsEngine === 'parler') {
325
- return resolveParlerConfig(cfg);
394
+ return resolveParlerConfig(cfg, ctx);
326
395
  }
327
396
  if (ttsEngine === 'cosyvoice3') {
328
397
  return resolveCosyvoice3Config(cfg, ctx);
@@ -339,7 +408,7 @@ export const ttsPlugin = definePlugin({
339
408
  const config = (params.modelConfig ?? {});
340
409
  const artifacts = params.artifacts ?? {};
341
410
  if (config.ttsEngine === 'parler') {
342
- return createParlerModel(params.modelId, config, params);
411
+ return createParlerModel(params.modelId, config, params, artifacts);
343
412
  }
344
413
  if (config.ttsEngine === 'cosyvoice3') {
345
414
  return createCosyvoice3Model(params.modelId, config, params, artifacts);
@@ -367,21 +436,16 @@ export const ttsPlugin = definePlugin({
367
436
  type: 'textToSpeech',
368
437
  buffer: result.value.buffer,
369
438
  done: false,
370
- ...(result.value.chunkIndex !== undefined
371
- ? { chunkIndex: result.value.chunkIndex }
372
- : {}),
373
- ...(typeof result.value.sentenceChunk === 'string' &&
374
- result.value.sentenceChunk.length > 0
375
- ? { sentenceChunk: result.value.sentenceChunk }
376
- : {})
439
+ ...ttsFrameMetadata(result.value)
377
440
  };
378
441
  result = await stream.next();
379
442
  }
380
- const { modelExecutionMs, stats } = result.value;
443
+ const { modelExecutionMs, stats, cancelled } = result.value;
381
444
  yield attachModelExecutionMs({
382
445
  type: 'textToSpeech',
383
446
  buffer: [],
384
447
  done: true,
448
+ stopReason: cancelled ? 'cancelled' : 'completed',
385
449
  ...(stats && { stats })
386
450
  }, modelExecutionMs);
387
451
  }
@@ -405,21 +469,16 @@ export const ttsPlugin = definePlugin({
405
469
  type: 'textToSpeechStream',
406
470
  buffer: result.value.buffer,
407
471
  done: false,
408
- ...(result.value.chunkIndex !== undefined
409
- ? { chunkIndex: result.value.chunkIndex }
410
- : {}),
411
- ...(typeof result.value.sentenceChunk === 'string' &&
412
- result.value.sentenceChunk.length > 0
413
- ? { sentenceChunk: result.value.sentenceChunk }
414
- : {})
472
+ ...ttsFrameMetadata(result.value)
415
473
  };
416
474
  result = await stream.next();
417
475
  }
418
- const { modelExecutionMs, stats } = result.value;
476
+ const { modelExecutionMs, stats, cancelled } = result.value;
419
477
  yield attachModelExecutionMs({
420
478
  type: 'textToSpeechStream',
421
479
  buffer: [],
422
480
  done: true,
481
+ stopReason: cancelled ? 'cancelled' : 'completed',
423
482
  ...(stats && { stats })
424
483
  }, modelExecutionMs);
425
484
  }
@@ -430,7 +489,7 @@ export const ttsPlugin = definePlugin({
430
489
  })
431
490
  },
432
491
  logging: {
433
- module: ttsAddonLogging,
492
+ module: () => import('@qvac/tts-ggml/addonLogging'),
434
493
  namespace: ModelType.ttsGgml
435
494
  }
436
495
  });
@@ -340,7 +340,7 @@ export declare const whisperPlugin: {
340
340
  }, import("zod/v4/core").$strip>>;
341
341
  };
342
342
  logging: {
343
- module: import("@qvac/asr-ggml/addonLogging").AddonLogging;
343
+ module: () => Promise<typeof import("@qvac/asr-ggml/addonLogging")>;
344
344
  namespace: "@qvac/asr-ggml";
345
345
  };
346
346
  };
@@ -1,4 +1,3 @@
1
- import asrAddonLogging from '@qvac/asr-ggml/addonLogging';
2
1
  import ASRGgml from '@qvac/asr-ggml';
3
2
  import { definePlugin, defineHandler, defineDuplexHandler, transcribeRequestSchema, transcribeResponseSchema, transcribeStreamRequestSchema, transcribeStreamResponseSchema, ModelType, whisperConfigSchema, ADDON_ASR } from '../../../schemas/index.js';
4
3
  import { transcribe, transcribeStream } from '../../ops/transcribe.js';
@@ -172,7 +171,7 @@ export const whisperPlugin = definePlugin({
172
171
  })
173
172
  },
174
173
  logging: {
175
- module: asrAddonLogging,
174
+ module: () => import('@qvac/asr-ggml/addonLogging'),
176
175
  namespace: ADDON_ASR
177
176
  }
178
177
  });
@@ -1,6 +1,6 @@
1
1
  import * as hostApi from '../api/index.js';
2
2
  import type { QvacPlugin } from '../schemas/plugin.js';
3
- export { registerPlugin, registerPlugins, getPlugin, getPluginHandler, getTurboVecIndexProvider, hasPlugin, getAllPlugins, clearPlugins, unregisterPlugin } from './registry.js';
3
+ export { registerPlugin, registerPlugins, ensureAddonLoggerReady, getPlugin, getPluginHandler, getTurboVecIndexProvider, hasPlugin, getAllPlugins, clearPlugins, unregisterPlugin } from './registry.js';
4
4
  /**
5
5
  * Register a set of plugins and return the host API bound to them. The
6
6
  * ergonomic entry point for assembling an explicit plugin subset — we ship
@@ -1,6 +1,6 @@
1
1
  import { registerPlugins } from './registry.js';
2
2
  import * as hostApi from '../api/index.js';
3
- export { registerPlugin, registerPlugins, getPlugin, getPluginHandler, getTurboVecIndexProvider, hasPlugin, getAllPlugins, clearPlugins, unregisterPlugin } from './registry.js';
3
+ export { registerPlugin, registerPlugins, ensureAddonLoggerReady, getPlugin, getPluginHandler, getTurboVecIndexProvider, hasPlugin, getAllPlugins, clearPlugins, unregisterPlugin } from './registry.js';
4
4
  /**
5
5
  * Register a set of plugins and return the host API bound to them. The
6
6
  * ergonomic entry point for assembling an explicit plugin subset — we ship
@@ -5,7 +5,8 @@ import { isModelLoaded, registerModel } from '../../runtime/model-registry.js';
5
5
  import { startLogBuffering, stopLogBufferingWithTimeout } from '../../runtime/logging-stream-registry.js';
6
6
  import { detectShardedModel, generateShardFilenames, validateShardedModelCache } from '../../utils/index.js';
7
7
  import { PluginNotFoundError, ModelFileNotFoundError, ModelFileNotFoundInDirError, ModelFileLocateFailedError } from '../../errors/index.js';
8
- import { getPlugin } from '../index.js';
8
+ import { ensureAddonLoggerReady, getPlugin } from '../index.js';
9
+ import { runAdvisoryFitCheck } from '../../resources/model-fit/native-probe/advisory-fit.js';
9
10
  import { promises as fsPromises } from 'bare-fs';
10
11
  import path from 'bare-path';
11
12
  import { getEngineLogger } from '../../logging/index.js';
@@ -52,6 +53,24 @@ export async function loadModel(params, options) {
52
53
  throw new ModelFileLocateFailedError(modelType, modelPath, error);
53
54
  }
54
55
  }
56
+ // Advisory: every outcome — including a projected insufficiency — continues
57
+ // to the ordinary load below. Runs after config resolution and path
58
+ // validation so it sees the same state the real load uses, and before
59
+ // `createModel()` so it never competes with the native load for device
60
+ // memory. The outcome is stored on the registry entry for internal use;
61
+ // it is deliberately not exposed on any public API yet.
62
+ const fitProbe = await runAdvisoryFitCheck({
63
+ modelId,
64
+ modelType: modelType,
65
+ modelPath,
66
+ modelConfig,
67
+ artifacts,
68
+ isShardedModel
69
+ });
70
+ // Load the addon's logger before the addon itself runs. Plugins that defer
71
+ // their logging module do so to keep registration free of native loads, so
72
+ // this is the first point their addon is genuinely required.
73
+ await ensureAddonLoggerReady(plugin);
55
74
  logger.info(`${modelType}: Loading model ${modelId}...`);
56
75
  startLogBuffering(modelId);
57
76
  try {
@@ -71,7 +90,8 @@ export async function loadModel(params, options) {
71
90
  path: modelPath,
72
91
  config: modelConfig,
73
92
  modelType: modelType,
74
- name: modelName
93
+ name: modelName,
94
+ fitProbe
75
95
  });
76
96
  const loadResult = modelInitializationTimeMs !== undefined ? { timing: { modelInitializationTimeMs } } : {};
77
97
  return forwardBackendDiagnostics(loadResult, result);
@@ -62,12 +62,50 @@ import { isAddonContextOverflowError } from '../builtin/llamacpp-completion/ops/
62
62
  import { generateRandomRequestId } from '../../runtime/request-id.js';
63
63
  import { getModelParallel } from '../../utils/config-transform.js';
64
64
  import { getEngineLogger } from '../../logging/index.js';
65
+ import { isTranslatePsyModel, buildTranslatePsyMessages, TRANSLATEPSY_GENERATION_PARAMS } from './translatepsy-utils.js';
66
+ const AFRICAN_LANGUAGE_ALIASES = new Map([
67
+ ['af', 'afr_Latn'],
68
+ ['sw', 'swh_Latn'],
69
+ ['so', 'som_Latn'],
70
+ ['am', 'amh_Ethi'],
71
+ ['ha', 'hau_Latn'],
72
+ ['rw', 'kin_Latn'],
73
+ ['ln', 'lin_Latn'],
74
+ ['lg', 'lug_Latn'],
75
+ ['zu', 'zul_Latn'],
76
+ ['ig', 'ibo_Latn'],
77
+ ['mg', 'plt_Latn'],
78
+ ['xh', 'xho_Latn'],
79
+ ['sn', 'sna_Latn'],
80
+ ['yo', 'yor_Latn'],
81
+ ['wo', 'wol_Latn'],
82
+ ['ny', 'nya_Latn'],
83
+ ['st', 'sot_Latn'],
84
+ ['ti', 'tir_Ethi'],
85
+ ['om', 'gaz_Latn'],
86
+ ['tn', 'tsn_Latn'],
87
+ ['swa', 'swh_Latn'],
88
+ ['mlg', 'plt_Latn'],
89
+ ['orm', 'gaz_Latn']
90
+ ]);
91
+ // Names missing from langdetect-text, kept separate from Afrique prompt selection.
92
+ const ADDITIONAL_LANGUAGE_NAMES = new Map([
93
+ ['lin', 'Lingala'],
94
+ ['lug', 'Luganda'],
95
+ ['wol', 'Wolof']
96
+ ]);
65
97
  export function getLanguage(code) {
66
98
  if (!code)
67
99
  return '';
68
- if (AFRICAN_LANGUAGES_MAP.has(code))
69
- return AFRICAN_LANGUAGES_MAP.get(code);
70
- const fullName = getLangName(code);
100
+ const languageCode = AFRICAN_LANGUAGE_ALIASES.get(code) ?? code;
101
+ if (AFRICAN_LANGUAGES_MAP.has(languageCode))
102
+ return AFRICAN_LANGUAGES_MAP.get(languageCode);
103
+ for (const [scriptCode, name] of AFRICAN_LANGUAGES_MAP) {
104
+ if (scriptCode.split('_')[0] === languageCode)
105
+ return name;
106
+ }
107
+ const baseCode = languageCode.split('_')[0];
108
+ const fullName = ADDITIONAL_LANGUAGE_NAMES.get(baseCode) ?? getLangName(baseCode);
71
109
  return fullName ?? code.toUpperCase();
72
110
  }
73
111
  export function isAfrican(code) {
@@ -119,6 +157,7 @@ export async function* translate(params, requestId) {
119
157
  from = detected.code;
120
158
  }
121
159
  const afriquePrompt = isLlm && (isAfrican(from) || isAfrican(to));
160
+ const translatePsy = isLlm && isTranslatePsyModel(entry.local);
122
161
  const fromLanguage = getLanguage(from);
123
162
  const toLanguage = getLanguage(to);
124
163
  // Open a request-scoped lifecycle for both engine branches. The LLM
@@ -184,14 +223,16 @@ export async function* translate(params, requestId) {
184
223
  // Prepare input based on model type
185
224
  const input = canonicalModelType === ModelType.nmtcppTranslation
186
225
  ? singleText
187
- : [
188
- {
189
- role: afriquePrompt ? 'user' : 'system',
190
- content: afriquePrompt
191
- ? `Translate ${fromLanguage} to ${toLanguage}.\n${fromLanguage}: ${singleText}\n${toLanguage}:`
192
- : `${context ? `${context}. ` : ''}Translate the following text from ${fromLanguage} into ${toLanguage}. Only output the translation, nothing else.\n\n${fromLanguage}: ${singleText}\n${toLanguage}:`
193
- }
194
- ];
226
+ : translatePsy
227
+ ? buildTranslatePsyMessages(fromLanguage, toLanguage, singleText)
228
+ : [
229
+ {
230
+ role: afriquePrompt ? 'user' : 'system',
231
+ content: afriquePrompt
232
+ ? `Translate ${fromLanguage} to ${toLanguage}.\n${fromLanguage}: ${singleText}\n${toLanguage}:`
233
+ : `${context ? `${context}. ` : ''}Translate the following text from ${fromLanguage} into ${toLanguage}. Only output the translation, nothing else.\n\n${fromLanguage}: ${singleText}\n${toLanguage}:`
234
+ }
235
+ ];
195
236
  // A translate cancelled before native work (queued-cancel, or the model being
196
237
  // unloaded) must not call run(): LLM could decode against an exclusive finetune
197
238
  // holding the lane, and either engine would run against a model being torn
@@ -203,7 +244,12 @@ export async function* translate(params, requestId) {
203
244
  }
204
245
  const modelStart = nowMs();
205
246
  let response;
206
- if (canonicalModelType === ModelType.llamacppCompletion &&
247
+ if (translatePsy) {
248
+ response = await model.run(input, {
249
+ generationParams: TRANSLATEPSY_GENERATION_PARAMS
250
+ });
251
+ }
252
+ else if (canonicalModelType === ModelType.llamacppCompletion &&
207
253
  !shouldSkipPerCallSampling(entry.local.name)) {
208
254
  response = await model.run(input, {
209
255
  generationParams: LLM_TRANSLATE_GENERATION_PARAMS
@@ -0,0 +1,17 @@
1
+ export declare function isTranslatePsyModel(model: {
2
+ name?: string | undefined;
3
+ path: string;
4
+ }): boolean;
5
+ export declare const TRANSLATEPSY_GENERATION_PARAMS: {
6
+ temp: number;
7
+ top_k: number;
8
+ top_p: number;
9
+ repeat_penalty: number;
10
+ predict: number;
11
+ seed: number;
12
+ reasoning_budget: number;
13
+ };
14
+ export declare function buildTranslatePsyMessages(from: string, to: string, text: string): {
15
+ role: string;
16
+ content: string;
17
+ }[];
@@ -0,0 +1,32 @@
1
+ export function isTranslatePsyModel(model) {
2
+ const filename = model.path.split(/[/\\]/).pop() ?? '';
3
+ return (/^TRANSLATEPSY_AFRISLM_/.test(model.name ?? '') ||
4
+ /^(?:[a-f0-9]{16}_)?TranslatePsy-AfriSLM-(?:0\.8|2|4)B-(?:Q4_K_M|Q8_0)-imat\.gguf$/i.test(filename));
5
+ }
6
+ export const TRANSLATEPSY_GENERATION_PARAMS = {
7
+ temp: 0,
8
+ top_k: 1,
9
+ top_p: 1,
10
+ repeat_penalty: 1,
11
+ predict: 256,
12
+ seed: 0,
13
+ reasoning_budget: 0
14
+ };
15
+ export function buildTranslatePsyMessages(from, to, text) {
16
+ // Match the model card exactly.
17
+ return [
18
+ {
19
+ role: 'system',
20
+ content: `You are a professional ${from} to ${to} translator. ` +
21
+ `Your goal is to accurately convey the meaning and nuances of the ` +
22
+ `original ${from} text while adhering to ${to} grammar, ` +
23
+ `vocabulary, and cultural sensitivities. Produce only the ` +
24
+ `${to} translation, without any additional explanations ` +
25
+ 'or commentary. '
26
+ },
27
+ {
28
+ role: 'user',
29
+ content: `Please translate the following ${from} text into ${to}: ${text}.\n\nTranslation:`
30
+ }
31
+ ];
32
+ }
@@ -1,6 +1,21 @@
1
1
  import { type QvacPlugin, type PluginHandlerDefinition } from '../schemas/plugin.js';
2
2
  export declare function registerPlugin(plugin: QvacPlugin): void;
3
3
  export declare function registerPlugins(pluginList: readonly QvacPlugin[]): void;
4
+ /**
5
+ * Wires the addon logger of a plugin that supplied a resolver, loading the
6
+ * addon on the way. Call it once a model of this type is actually being
7
+ * created: it is the point where the addon is needed anyway, so a host
8
+ * missing that addon's platform package fails the load it asked for instead
9
+ * of failing every plugin registration at startup.
10
+ *
11
+ * A no-op for plugins that passed their module directly — those are already
12
+ * wired by `registerPlugin` — and for a namespace that is already wired,
13
+ * which is how two plugins over one addon (whisper and Parakeet over ASR)
14
+ * share a single `setLogger` call. Also a no-op when the registry was cleared
15
+ * while the addon was loading, so a shutdown that races a load leaves no
16
+ * logger behind.
17
+ */
18
+ export declare function ensureAddonLoggerReady(plugin: QvacPlugin): Promise<void>;
4
19
  export declare function getPlugin(modelType: string): QvacPlugin | undefined;
5
20
  export declare function getPluginHandler(modelType: string, handlerName: string): PluginHandlerDefinition | undefined;
6
21
  export declare function hasPlugin(modelType: string): boolean;