@qvac/inference 0.20.0 → 0.21.0

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 (213) hide show
  1. package/NOTICE +107 -33
  2. package/README.md +8 -5
  3. package/dist/api/assess-model-fit.d.ts +9 -4
  4. package/dist/api/assess-model-fit.js +9 -4
  5. package/dist/api/bci-transcribe.d.ts +4 -1
  6. package/dist/api/index.d.ts +1 -0
  7. package/dist/api/index.js +1 -0
  8. package/dist/api/rpc-server.d.ts +19 -0
  9. package/dist/api/rpc-server.js +44 -0
  10. package/dist/api/text-to-speech.d.ts +9 -0
  11. package/dist/api/text-to-speech.js +9 -0
  12. package/dist/api/transcribe.d.ts +11 -7
  13. package/dist/api/transcribe.js +47 -15
  14. package/dist/api/video-request.d.ts +5 -5
  15. package/dist/dispatch.js +39 -7
  16. package/dist/errors/index.d.ts +7 -0
  17. package/dist/errors/index.js +15 -0
  18. package/dist/handlers/assess-model-fit.d.ts +18 -6
  19. package/dist/handlers/assess-model-fit.js +215 -35
  20. package/dist/handlers/completion-orchestrate.js +17 -0
  21. package/dist/handlers/get-loaded-model-info.js +4 -1
  22. package/dist/handlers/get-model-info.js +29 -1
  23. package/dist/handlers/load-model/coreml-sidecars.d.ts +25 -0
  24. package/dist/handlers/load-model/coreml-sidecars.js +274 -0
  25. package/dist/handlers/load-model/hyperdrive.js +2 -2
  26. package/dist/handlers/load-model/registry-download-utils.d.ts +1 -0
  27. package/dist/handlers/load-model/registry-download-utils.js +2 -1
  28. package/dist/handlers/load-model/registry.js +59 -0
  29. package/dist/handlers/rpc-server.d.ts +23 -0
  30. package/dist/handlers/rpc-server.js +141 -0
  31. package/dist/index.d.ts +2 -1
  32. package/dist/index.js +2 -1
  33. package/dist/models/registry/models.d.ts +7731 -5672
  34. package/dist/models/registry/models.js +4652 -2526
  35. package/dist/models/registry/resource-profiles.js +481 -94
  36. package/dist/plugins/builtin/audiogen-ggml/plugin.d.ts +8 -6
  37. package/dist/plugins/builtin/audiogen-ggml/plugin.js +2 -1
  38. package/dist/plugins/builtin/bci-whispercpp-transcription/plugin.d.ts +73 -1
  39. package/dist/plugins/builtin/bci-whispercpp-transcription/plugin.js +37 -19
  40. package/dist/plugins/builtin/ggml-vla/plugin.d.ts +1 -1
  41. package/dist/plugins/builtin/llamacpp-completion/ops/batch-completion-stream.js +9 -5
  42. package/dist/plugins/builtin/llamacpp-completion/ops/completion-stats.d.ts +1 -1
  43. package/dist/plugins/builtin/llamacpp-completion/ops/completion-stream.js +16 -7
  44. package/dist/plugins/builtin/llamacpp-completion/plugin.d.ts +80 -8
  45. package/dist/plugins/builtin/llamacpp-completion/plugin.js +9 -3
  46. package/dist/plugins/builtin/llamacpp-completion/transform.js +19 -0
  47. package/dist/plugins/builtin/llamacpp-embedding/plugin.d.ts +5 -3
  48. package/dist/plugins/builtin/llamacpp-embedding/plugin.js +2 -1
  49. package/dist/plugins/builtin/parakeet-transcription/plugin.d.ts +121 -0
  50. package/dist/plugins/builtin/parakeet-transcription/plugin.js +56 -27
  51. package/dist/plugins/builtin/sdcpp-generation/plugin.d.ts +8 -2
  52. package/dist/plugins/builtin/sdcpp-generation/plugin.js +7 -2
  53. package/dist/plugins/builtin/tts-ggml/ops/parler-options.d.ts +2 -0
  54. package/dist/plugins/builtin/tts-ggml/ops/parler-options.js +5 -1
  55. package/dist/plugins/builtin/tts-ggml/ops/text-to-speech.js +27 -10
  56. package/dist/plugins/builtin/tts-ggml/plugin.d.ts +181 -0
  57. package/dist/plugins/builtin/tts-ggml/plugin.js +84 -2
  58. package/dist/plugins/builtin/whispercpp-transcription/plugin.d.ts +136 -0
  59. package/dist/plugins/builtin/whispercpp-transcription/plugin.js +18 -7
  60. package/dist/plugins/ops/bci-transcribe.d.ts +6 -2
  61. package/dist/plugins/ops/bci-transcribe.js +18 -13
  62. package/dist/plugins/ops/load-model.js +2 -2
  63. package/dist/plugins/ops/transcribe.d.ts +4 -2
  64. package/dist/plugins/ops/transcribe.js +39 -2
  65. package/dist/plugins/ops/translate-stats.d.ts +2 -1
  66. package/dist/plugins/ops/translate-stats.js +64 -7
  67. package/dist/plugins/ops/translate.js +20 -4
  68. package/dist/profiling/operation-metrics.js +56 -0
  69. package/dist/registry.js +4 -0
  70. package/dist/resources/model-fit/assess.d.ts +46 -2
  71. package/dist/resources/model-fit/assess.js +67 -3
  72. package/dist/resources/model-fit/estimators/llm.js +2 -2
  73. package/dist/resources/model-fit/estimators/whisper.js +3 -4
  74. package/dist/resources/model-fit/fit-stub/fetch-fit-stub.d.ts +102 -0
  75. package/dist/resources/model-fit/fit-stub/fetch-fit-stub.js +212 -0
  76. package/dist/resources/model-fit/fit-stub/project-fit-from-load.d.ts +27 -0
  77. package/dist/resources/model-fit/fit-stub/project-fit-from-load.js +31 -0
  78. package/dist/resources/model-fit/fit-stub/resolve-load-from-stubs.d.ts +38 -0
  79. package/dist/resources/model-fit/fit-stub/resolve-load-from-stubs.js +138 -0
  80. package/dist/resources/model-fit/floor.js +2 -2
  81. package/dist/resources/model-fit/native-probe/advisory-fit.d.ts +9 -6
  82. package/dist/resources/model-fit/native-probe/advisory-fit.js +129 -80
  83. package/dist/resources/model-fit/native-probe/child-engine-fit.d.ts +2 -0
  84. package/dist/resources/model-fit/native-probe/child-engine-fit.js +41 -0
  85. package/dist/resources/model-fit/native-probe/classify-fit.d.ts +6 -0
  86. package/dist/resources/model-fit/native-probe/classify-fit.js +135 -0
  87. package/dist/resources/model-fit/native-probe/create-fit-request.d.ts +28 -0
  88. package/dist/resources/model-fit/native-probe/create-fit-request.js +51 -0
  89. package/dist/resources/model-fit/native-probe/engine-fit.d.ts +59 -0
  90. package/dist/resources/model-fit/native-probe/engine-fit.js +13 -0
  91. package/dist/resources/model-fit/native-probe/engines/asr.d.ts +10 -0
  92. package/dist/resources/model-fit/native-probe/engines/asr.js +103 -0
  93. package/dist/resources/model-fit/native-probe/engines/audiogen.d.ts +8 -0
  94. package/dist/resources/model-fit/native-probe/engines/audiogen.js +35 -0
  95. package/dist/resources/model-fit/native-probe/engines/diffusion.d.ts +7 -0
  96. package/dist/resources/model-fit/native-probe/engines/diffusion.js +55 -0
  97. package/dist/resources/model-fit/native-probe/engines/gpu.d.ts +10 -0
  98. package/dist/resources/model-fit/native-probe/engines/gpu.js +21 -0
  99. package/dist/resources/model-fit/native-probe/engines/llama.d.ts +13 -0
  100. package/dist/resources/model-fit/native-probe/engines/llama.js +145 -0
  101. package/dist/resources/model-fit/native-probe/engines/tts.d.ts +8 -0
  102. package/dist/resources/model-fit/native-probe/engines/tts.js +106 -0
  103. package/dist/resources/model-fit/native-probe/fit-outcome.d.ts +16 -0
  104. package/dist/resources/model-fit/native-probe/fit-outcome.js +1 -0
  105. package/dist/resources/model-fit/native-probe/fit-process.d.ts +27 -0
  106. package/dist/resources/model-fit/native-probe/fit-process.js +55 -0
  107. package/dist/resources/model-fit/native-probe/fit-runner.d.ts +1 -0
  108. package/dist/resources/model-fit/native-probe/fit-runner.js +58 -0
  109. package/dist/resources/model-fit/native-probe/run-fit.d.ts +24 -0
  110. package/dist/resources/model-fit/native-probe/run-fit.js +20 -0
  111. package/dist/resources/model-fit/native-probe/run-in-process-fit.d.ts +14 -34
  112. package/dist/resources/model-fit/native-probe/run-in-process-fit.js +55 -139
  113. package/dist/resources/model-fit/native-probe/run-isolated-fit.d.ts +5 -12
  114. package/dist/resources/model-fit/native-probe/run-isolated-fit.js +13 -8
  115. package/dist/resources/model-fit/platform.d.ts +9 -0
  116. package/dist/resources/model-fit/platform.js +28 -0
  117. package/dist/rpc/device-map.d.ts +13 -0
  118. package/dist/rpc/device-map.js +22 -0
  119. package/dist/rpc/discovery.d.ts +12 -0
  120. package/dist/rpc/discovery.js +225 -0
  121. package/dist/rpc/ggml-provider.d.ts +4 -0
  122. package/dist/rpc/ggml-provider.js +25 -0
  123. package/dist/rpc/index.d.ts +2 -0
  124. package/dist/rpc/index.js +1 -0
  125. package/dist/rpc/instance.d.ts +6 -0
  126. package/dist/rpc/instance.js +21 -0
  127. package/dist/rpc/inventory.d.ts +5 -0
  128. package/dist/rpc/inventory.js +144 -0
  129. package/dist/rpc/manager.d.ts +14 -0
  130. package/dist/rpc/manager.js +174 -0
  131. package/dist/rpc/network.d.ts +11 -0
  132. package/dist/rpc/network.js +85 -0
  133. package/dist/rpc/provider.d.ts +6 -0
  134. package/dist/rpc/provider.js +24 -0
  135. package/dist/runtime/lifecycle.js +31 -13
  136. package/dist/runtime/request-context.d.ts +1 -1
  137. package/dist/runtime/state.js +5 -0
  138. package/dist/schemas/assess-model-fit.d.ts +284 -77
  139. package/dist/schemas/assess-model-fit.js +93 -55
  140. package/dist/schemas/audio-gen.d.ts +6 -6
  141. package/dist/schemas/audio-gen.js +2 -2
  142. package/dist/schemas/batch-completion-stream.d.ts +13 -5
  143. package/dist/schemas/batch-completion-stream.js +5 -3
  144. package/dist/schemas/bci-config.js +1 -1
  145. package/dist/schemas/bci.d.ts +86 -1
  146. package/dist/schemas/bci.js +21 -2
  147. package/dist/schemas/cancel.d.ts +10 -0
  148. package/dist/schemas/cancel.js +2 -0
  149. package/dist/schemas/common.d.ts +765 -56
  150. package/dist/schemas/common.js +7 -0
  151. package/dist/schemas/completion-event.d.ts +6 -6
  152. package/dist/schemas/completion-stream.d.ts +27 -8
  153. package/dist/schemas/completion-stream.js +41 -4
  154. package/dist/schemas/config.d.ts +236 -30
  155. package/dist/schemas/config.js +10 -0
  156. package/dist/schemas/embed.d.ts +2 -2
  157. package/dist/schemas/error.js +13 -0
  158. package/dist/schemas/errors.d.ts +1 -0
  159. package/dist/schemas/errors.js +5 -0
  160. package/dist/schemas/index.d.ts +1 -0
  161. package/dist/schemas/index.js +1 -0
  162. package/dist/schemas/llamacpp-config.d.ts +105 -14
  163. package/dist/schemas/llamacpp-config.js +113 -3
  164. package/dist/schemas/load-model.d.ts +3102 -627
  165. package/dist/schemas/load-model.js +2 -2
  166. package/dist/schemas/mcp-adapter.d.ts +4 -0
  167. package/dist/schemas/model-config-schemas.d.ts +205 -5
  168. package/dist/schemas/model-info.d.ts +56 -0
  169. package/dist/schemas/model-info.js +5 -1
  170. package/dist/schemas/plugin.d.ts +8 -2
  171. package/dist/schemas/profiling.d.ts +4 -4
  172. package/dist/schemas/reload-config.d.ts +30 -0
  173. package/dist/schemas/rpc-server.d.ts +101 -0
  174. package/dist/schemas/rpc-server.js +87 -0
  175. package/dist/schemas/sdcpp-config.d.ts +7 -3
  176. package/dist/schemas/sdcpp-config.js +37 -2
  177. package/dist/schemas/system-resources.d.ts +4 -4
  178. package/dist/schemas/text-to-speech.d.ts +529 -1
  179. package/dist/schemas/text-to-speech.js +122 -12
  180. package/dist/schemas/tools.d.ts +2 -0
  181. package/dist/schemas/tools.js +8 -0
  182. package/dist/schemas/transcription-config.d.ts +5 -0
  183. package/dist/schemas/transcription-config.js +20 -2
  184. package/dist/schemas/transcription.d.ts +164 -0
  185. package/dist/schemas/transcription.js +71 -4
  186. package/dist/schemas/vla.d.ts +1 -1
  187. package/dist/surface.d.ts +4 -1
  188. package/dist/surface.js +3 -1
  189. package/dist/utils/addon-responses.d.ts +5 -1
  190. package/dist/utils/archive.js +5 -1
  191. package/dist/utils/asr-diagnostics.d.ts +12 -0
  192. package/dist/utils/asr-diagnostics.js +42 -0
  193. package/dist/utils/asr-events.d.ts +1 -0
  194. package/dist/utils/asr-events.js +3 -2
  195. package/dist/utils/cache/paths.js +3 -2
  196. package/dist/utils/mcp-adapter.js +7 -1
  197. package/dist/utils/multi-gpu-mobile.d.ts +1 -1
  198. package/dist/utils/multi-gpu-mobile.js +8 -6
  199. package/dist/utils/path-security.d.ts +5 -0
  200. package/dist/utils/path-security.js +14 -0
  201. package/dist/utils/tool-helpers.d.ts +4 -0
  202. package/dist/utils/tool-helpers.js +2 -0
  203. package/dist/utils/tools/defer.d.ts +93 -0
  204. package/dist/utils/tools/defer.js +290 -0
  205. package/dist/utils/tools/index.d.ts +1 -0
  206. package/dist/utils/tools/index.js +1 -0
  207. package/dist/utils/transcribe-metadata.d.ts +21 -7
  208. package/dist/utils/transcribe-metadata.js +24 -8
  209. package/dist/utils/tts-stats.d.ts +6 -0
  210. package/dist/utils/tts-stats.js +14 -1
  211. package/package.json +44 -30
  212. package/dist/resources/model-fit/native-probe/create-llama-fit-request.d.ts +0 -26
  213. package/dist/resources/model-fit/native-probe/create-llama-fit-request.js +0 -171
@@ -71,14 +71,15 @@ import { isAddonContextOverflowError, parseContextOverflowMessage } from './ops/
71
71
  import { stoppedByLength } from './ops/completion-stats.js';
72
72
  import { isMobile } from '../../../runtime/state.js';
73
73
  import { stripMultiGpuKeys } from '../../../utils/multi-gpu-mobile.js';
74
+ import { resolveDeferredTools, toWireTool } from '../../../utils/tools/defer.js';
74
75
  function createLlmModel(modelId, modelPath, llmConfig, projectionModelPath) {
75
76
  const logger = createStreamLogger(modelId, ModelType.llamacppCompletion);
76
77
  registerAddonLogger(modelId, ModelType.llamacppCompletion, logger);
77
78
  const llmConfigStrings = transformLlmConfig(llmConfig);
78
79
  if (isMobile()) {
79
- const stripped = stripMultiGpuKeys(llmConfigStrings);
80
+ const stripped = stripMultiGpuKeys(llmConfigStrings, true);
80
81
  if (stripped.length > 0) {
81
- getEngineLogger().warn(`[${ModelType.llamacppCompletion}:${modelId}] Multi-GPU parameters (${stripped.join(', ')}) are not supported on mobile (single-GPU device) — removing from config; model will load with single-GPU defaults`);
82
+ getEngineLogger().warn(`[${ModelType.llamacppCompletion}:${modelId}] Removing unsupported mobile parameters: ${stripped.join(', ')}`);
82
83
  }
83
84
  }
84
85
  // Addon overflow flag: at a single slot the addon throws when driven over
@@ -116,6 +117,7 @@ export const llmPlugin = definePlugin({
116
117
  displayName: 'LLM (llama.cpp)',
117
118
  addonPackage: ADDON_LLM,
118
119
  loadConfigSchema: llmConfigBaseSchema,
120
+ assessFit: LlmLlamacpp.assessFit,
119
121
  async resolveConfig(cfg, ctx) {
120
122
  const { projectionModelSrc, ...llmConfig } = cfg;
121
123
  if (!projectionModelSrc) {
@@ -145,7 +147,11 @@ export const llmPlugin = definePlugin({
145
147
  const dialect = request.toolDialect ?? detectToolDialect(request.modelId);
146
148
  const modelCfg = getModelConfig(request.modelId);
147
149
  const toolsActive = modelCfg.tools === true;
148
- const toolsByPosition = request.prompts.map((prompt) => prompt.tools ?? []);
150
+ // Parse against what each prompt can call: with deferred tools that is
151
+ // the rendered set plus whatever an earlier `tool_search` loaded.
152
+ const toolsByPosition = request.prompts.map((prompt) => resolveDeferredTools(prompt.tools, prompt.history)?.callableTools ??
153
+ prompt.tools?.map(toWireTool) ??
154
+ []);
149
155
  const toolsById = new Map();
150
156
  const normalizers = new Map();
151
157
  const bufferedEvents = [];
@@ -23,5 +23,24 @@ export function transformLlmConfig(llmConfig) {
23
23
  transformed['openclCacheDir'] = transformed['opencl_cache_dir'];
24
24
  delete transformed['opencl_cache_dir'];
25
25
  }
26
+ if ('cpu-moe' in transformed) {
27
+ if (transformed['cpu-moe'] === 'true') {
28
+ transformed['cpu-moe'] = '';
29
+ }
30
+ else {
31
+ delete transformed['cpu-moe'];
32
+ }
33
+ }
34
+ if ('kv-offload' in transformed) {
35
+ const enabled = transformed['kv-offload'] === 'true';
36
+ delete transformed['kv-offload'];
37
+ transformed[enabled ? 'kv-offload' : 'no-kv-offload'] = '';
38
+ }
39
+ if (transformed['prefetch-weights'] === 'true') {
40
+ transformed['prefetch-weights'] = '1';
41
+ }
42
+ else if (transformed['prefetch-weights'] === 'false') {
43
+ transformed['prefetch-weights'] = '0';
44
+ }
26
45
  return transformed;
27
46
  }
@@ -1,3 +1,4 @@
1
+ import { assessFit as embedAssessFit } from '@qvac/embed-llamacpp';
1
2
  import type { TurboVecIndexProvider } from '@qvac/rag';
2
3
  import { type CreateModelParams, type PluginModelResult } from '../../../schemas/index.js';
3
4
  export declare const embeddingsPlugin: {
@@ -7,8 +8,8 @@ export declare const embeddingsPlugin: {
7
8
  loadConfigSchema: import("zod").ZodObject<{
8
9
  gpuLayers: import("zod").ZodOptional<import("zod").ZodNumber>;
9
10
  device: import("zod").ZodOptional<import("zod").ZodEnum<{
10
- gpu: "gpu";
11
11
  cpu: "cpu";
12
+ gpu: "gpu";
12
13
  }>>;
13
14
  batchSize: import("zod").ZodOptional<import("zod").ZodNumber>;
14
15
  pooling: import("zod").ZodOptional<import("zod").ZodEnum<{
@@ -24,9 +25,9 @@ export declare const embeddingsPlugin: {
24
25
  }>>;
25
26
  embdNormalize: import("zod").ZodOptional<import("zod").ZodNumber>;
26
27
  flashAttention: import("zod").ZodOptional<import("zod").ZodEnum<{
28
+ auto: "auto";
27
29
  on: "on";
28
30
  off: "off";
29
- auto: "auto";
30
31
  }>>;
31
32
  mainGpu: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodNumber, import("zod").ZodEnum<{
32
33
  integrated: "integrated";
@@ -45,6 +46,7 @@ export declare const embeddingsPlugin: {
45
46
  }>>;
46
47
  openclCacheDir: import("zod").ZodOptional<import("zod").ZodString>;
47
48
  }, import("zod/v4/core").$strip>;
49
+ assessFit: typeof embedAssessFit;
48
50
  createModel(params: CreateModelParams): PluginModelResult;
49
51
  handlers: {
50
52
  embed: import("../../../schemas/index.js").PluginHandlerDefinition<import("zod").ZodObject<{
@@ -61,8 +63,8 @@ export declare const embeddingsPlugin: {
61
63
  tokensPerSecond: import("zod").ZodOptional<import("zod").ZodNumber>;
62
64
  totalTokens: import("zod").ZodOptional<import("zod").ZodNumber>;
63
65
  backendDevice: import("zod").ZodOptional<import("zod").ZodEnum<{
64
- gpu: "gpu";
65
66
  cpu: "cpu";
67
+ gpu: "gpu";
66
68
  }>>;
67
69
  contextSize: import("zod").ZodOptional<import("zod").ZodNumber>;
68
70
  }, import("zod/v4/core").$strip>>;
@@ -1,4 +1,4 @@
1
- import EmbedLlamacpp, { IdMapIndex } from '@qvac/embed-llamacpp';
1
+ import EmbedLlamacpp, { IdMapIndex, assessFit as embedAssessFit } from '@qvac/embed-llamacpp';
2
2
  import { definePlugin, defineHandler, embedRequestSchema, embedResponseSchema, ModelType, embedConfigBaseSchema, ADDON_EMBEDDING } from '../../../schemas/index.js';
3
3
  import { createStreamLogger, registerAddonLogger, getEngineLogger } from '../../../logging/index.js';
4
4
  import { getFirstShardPath } from '../../../utils/index.js';
@@ -38,6 +38,7 @@ export const embeddingsPlugin = definePlugin({
38
38
  displayName: 'Embeddings (llama.cpp)',
39
39
  addonPackage: ADDON_EMBEDDING,
40
40
  loadConfigSchema: embedConfigBaseSchema,
41
+ assessFit: embedAssessFit,
41
42
  createModel(params) {
42
43
  const embedConfig = (params.modelConfig ?? {});
43
44
  const { model } = createEmbeddingsModel(params.modelId, params.modelPath, embedConfig);
@@ -6,6 +6,7 @@ export declare const parakeetPlugin: {
6
6
  loadConfigSchema: import("zod").ZodObject<{
7
7
  [x: string]: import("zod").ZodOptional<import("zod").ZodUnknown>;
8
8
  }, import("zod/v4/core").$strict>;
9
+ assessFit: typeof import("node_modules/@qvac/asr-ggml/lib/fit").assessFit;
9
10
  resolveConfig(cfg: ParakeetConfig): Promise<ResolveResult<ParakeetConfig>>;
10
11
  createModel(params: CreateModelParams): PluginModelResult;
11
12
  handlers: {
@@ -41,6 +42,18 @@ export declare const parakeetPlugin: {
41
42
  gpuUnsupported: import("zod").ZodOptional<import("zod").ZodNumber>;
42
43
  gpuMemTotalMb: import("zod").ZodOptional<import("zod").ZodNumber>;
43
44
  gpuMemFreeMb: import("zod").ZodOptional<import("zod").ZodNumber>;
45
+ totalTime: import("zod").ZodOptional<import("zod").ZodNumber>;
46
+ totalWallMs: import("zod").ZodOptional<import("zod").ZodNumber>;
47
+ totalSamples: import("zod").ZodOptional<import("zod").ZodNumber>;
48
+ processCalls: import("zod").ZodOptional<import("zod").ZodNumber>;
49
+ whisperSampleMs: import("zod").ZodOptional<import("zod").ZodNumber>;
50
+ whisperBatchdMs: import("zod").ZodOptional<import("zod").ZodNumber>;
51
+ whisperPromptMs: import("zod").ZodOptional<import("zod").ZodNumber>;
52
+ totalTranscriptions: import("zod").ZodOptional<import("zod").ZodNumber>;
53
+ modelLoadMs: import("zod").ZodOptional<import("zod").ZodNumber>;
54
+ totalEncodedFrames: import("zod").ZodOptional<import("zod").ZodNumber>;
55
+ encoderOnCoreml: import("zod").ZodOptional<import("zod").ZodNumber>;
56
+ encoderUsedCoreml: import("zod").ZodOptional<import("zod").ZodNumber>;
44
57
  }, import("zod/v4/core").$strip>>;
45
58
  error: import("zod").ZodOptional<import("zod").ZodString>;
46
59
  segment: import("zod").ZodOptional<import("zod").ZodObject<{
@@ -49,10 +62,17 @@ export declare const parakeetPlugin: {
49
62
  endMs: import("zod").ZodNumber;
50
63
  append: import("zod").ZodBoolean;
51
64
  id: import("zod").ZodNumber;
65
+ isEndOfTurn: import("zod").ZodOptional<import("zod").ZodBoolean>;
66
+ startsWord: import("zod").ZodOptional<import("zod").ZodBoolean>;
52
67
  }, import("zod/v4/core").$strip>>;
53
68
  vad: import("zod").ZodOptional<import("zod").ZodObject<{
54
69
  speaking: import("zod").ZodBoolean;
55
70
  probability: import("zod").ZodNumber;
71
+ source: import("zod").ZodOptional<import("zod").ZodEnum<{
72
+ silero: "silero";
73
+ energy: "energy";
74
+ sortformer: "sortformer";
75
+ }>>;
56
76
  }, import("zod/v4/core").$strip>>;
57
77
  endOfTurn: import("zod").ZodOptional<import("zod").ZodPreprocess<import("zod").ZodDiscriminatedUnion<[import("zod").ZodObject<{
58
78
  source: import("zod").ZodLiteral<"whisper">;
@@ -60,6 +80,47 @@ export declare const parakeetPlugin: {
60
80
  }, import("zod/v4/core").$strip>, import("zod").ZodObject<{
61
81
  source: import("zod").ZodLiteral<"parakeet">;
62
82
  }, import("zod/v4/core").$strip>], "source">, unknown>>;
83
+ diagnostics: import("zod").ZodOptional<import("zod").ZodObject<{
84
+ selectedBackend: import("zod").ZodString;
85
+ selectedDevice: import("zod").ZodEnum<{
86
+ cpu: "cpu";
87
+ gpu: "gpu";
88
+ }>;
89
+ graphicsApi: import("zod").ZodOptional<import("zod").ZodEnum<{
90
+ vulkan: "vulkan";
91
+ metal: "metal";
92
+ opencl: "opencl";
93
+ cuda: "cuda";
94
+ opengl: "opengl";
95
+ webgpu: "webgpu";
96
+ direct3d11: "direct3d11";
97
+ direct3d12: "direct3d12";
98
+ levelZero: "levelZero";
99
+ rocm: "rocm";
100
+ }>>;
101
+ driver: import("zod").ZodOptional<import("zod").ZodObject<{
102
+ name: import("zod").ZodString;
103
+ version: import("zod").ZodOptional<import("zod").ZodString>;
104
+ }, import("zod/v4/core").$strip>>;
105
+ gpuId: import("zod").ZodOptional<import("zod").ZodString>;
106
+ fallback: import("zod").ZodOptional<import("zod").ZodObject<{
107
+ requestedBackend: import("zod").ZodOptional<import("zod").ZodString>;
108
+ requestedDevice: import("zod").ZodOptional<import("zod").ZodEnum<{
109
+ cpu: "cpu";
110
+ gpu: "gpu";
111
+ }>>;
112
+ reason: import("zod").ZodString;
113
+ }, import("zod/v4/core").$strip>>;
114
+ probe: import("zod").ZodOptional<import("zod").ZodObject<{
115
+ status: import("zod").ZodEnum<{
116
+ unknown: "unknown";
117
+ compatible: "compatible";
118
+ incompatible: "incompatible";
119
+ }>;
120
+ backend: import("zod").ZodString;
121
+ reason: import("zod").ZodOptional<import("zod").ZodString>;
122
+ }, import("zod/v4/core").$strip>>;
123
+ }, import("zod/v4/core").$strip>>;
63
124
  type: import("zod").ZodLiteral<"transcribe">;
64
125
  }, import("zod/v4/core").$strip>>;
65
126
  transcribeStream: import("../../../schemas/index.js").PluginHandlerDefinition<import("zod").ZodObject<{
@@ -104,6 +165,18 @@ export declare const parakeetPlugin: {
104
165
  gpuUnsupported: import("zod").ZodOptional<import("zod").ZodNumber>;
105
166
  gpuMemTotalMb: import("zod").ZodOptional<import("zod").ZodNumber>;
106
167
  gpuMemFreeMb: import("zod").ZodOptional<import("zod").ZodNumber>;
168
+ totalTime: import("zod").ZodOptional<import("zod").ZodNumber>;
169
+ totalWallMs: import("zod").ZodOptional<import("zod").ZodNumber>;
170
+ totalSamples: import("zod").ZodOptional<import("zod").ZodNumber>;
171
+ processCalls: import("zod").ZodOptional<import("zod").ZodNumber>;
172
+ whisperSampleMs: import("zod").ZodOptional<import("zod").ZodNumber>;
173
+ whisperBatchdMs: import("zod").ZodOptional<import("zod").ZodNumber>;
174
+ whisperPromptMs: import("zod").ZodOptional<import("zod").ZodNumber>;
175
+ totalTranscriptions: import("zod").ZodOptional<import("zod").ZodNumber>;
176
+ modelLoadMs: import("zod").ZodOptional<import("zod").ZodNumber>;
177
+ totalEncodedFrames: import("zod").ZodOptional<import("zod").ZodNumber>;
178
+ encoderOnCoreml: import("zod").ZodOptional<import("zod").ZodNumber>;
179
+ encoderUsedCoreml: import("zod").ZodOptional<import("zod").ZodNumber>;
107
180
  }, import("zod/v4/core").$strip>>;
108
181
  error: import("zod").ZodOptional<import("zod").ZodString>;
109
182
  segment: import("zod").ZodOptional<import("zod").ZodObject<{
@@ -112,10 +185,17 @@ export declare const parakeetPlugin: {
112
185
  endMs: import("zod").ZodNumber;
113
186
  append: import("zod").ZodBoolean;
114
187
  id: import("zod").ZodNumber;
188
+ isEndOfTurn: import("zod").ZodOptional<import("zod").ZodBoolean>;
189
+ startsWord: import("zod").ZodOptional<import("zod").ZodBoolean>;
115
190
  }, import("zod/v4/core").$strip>>;
116
191
  vad: import("zod").ZodOptional<import("zod").ZodObject<{
117
192
  speaking: import("zod").ZodBoolean;
118
193
  probability: import("zod").ZodNumber;
194
+ source: import("zod").ZodOptional<import("zod").ZodEnum<{
195
+ silero: "silero";
196
+ energy: "energy";
197
+ sortformer: "sortformer";
198
+ }>>;
119
199
  }, import("zod/v4/core").$strip>>;
120
200
  endOfTurn: import("zod").ZodOptional<import("zod").ZodPreprocess<import("zod").ZodDiscriminatedUnion<[import("zod").ZodObject<{
121
201
  source: import("zod").ZodLiteral<"whisper">;
@@ -123,6 +203,47 @@ export declare const parakeetPlugin: {
123
203
  }, import("zod/v4/core").$strip>, import("zod").ZodObject<{
124
204
  source: import("zod").ZodLiteral<"parakeet">;
125
205
  }, import("zod/v4/core").$strip>], "source">, unknown>>;
206
+ diagnostics: import("zod").ZodOptional<import("zod").ZodObject<{
207
+ selectedBackend: import("zod").ZodString;
208
+ selectedDevice: import("zod").ZodEnum<{
209
+ cpu: "cpu";
210
+ gpu: "gpu";
211
+ }>;
212
+ graphicsApi: import("zod").ZodOptional<import("zod").ZodEnum<{
213
+ vulkan: "vulkan";
214
+ metal: "metal";
215
+ opencl: "opencl";
216
+ cuda: "cuda";
217
+ opengl: "opengl";
218
+ webgpu: "webgpu";
219
+ direct3d11: "direct3d11";
220
+ direct3d12: "direct3d12";
221
+ levelZero: "levelZero";
222
+ rocm: "rocm";
223
+ }>>;
224
+ driver: import("zod").ZodOptional<import("zod").ZodObject<{
225
+ name: import("zod").ZodString;
226
+ version: import("zod").ZodOptional<import("zod").ZodString>;
227
+ }, import("zod/v4/core").$strip>>;
228
+ gpuId: import("zod").ZodOptional<import("zod").ZodString>;
229
+ fallback: import("zod").ZodOptional<import("zod").ZodObject<{
230
+ requestedBackend: import("zod").ZodOptional<import("zod").ZodString>;
231
+ requestedDevice: import("zod").ZodOptional<import("zod").ZodEnum<{
232
+ cpu: "cpu";
233
+ gpu: "gpu";
234
+ }>>;
235
+ reason: import("zod").ZodString;
236
+ }, import("zod/v4/core").$strip>>;
237
+ probe: import("zod").ZodOptional<import("zod").ZodObject<{
238
+ status: import("zod").ZodEnum<{
239
+ unknown: "unknown";
240
+ compatible: "compatible";
241
+ incompatible: "incompatible";
242
+ }>;
243
+ backend: import("zod").ZodString;
244
+ reason: import("zod").ZodOptional<import("zod").ZodString>;
245
+ }, import("zod/v4/core").$strip>>;
246
+ }, import("zod/v4/core").$strip>>;
126
247
  type: import("zod").ZodLiteral<"transcribeStream">;
127
248
  }, import("zod/v4/core").$strip>>;
128
249
  };
@@ -1,8 +1,9 @@
1
1
  import ASRGgml from '@qvac/asr-ggml';
2
2
  import { definePlugin, defineHandler, defineDuplexHandler, transcribeRequestSchema, transcribeResponseSchema, transcribeStreamRequestSchema, transcribeStreamResponseSchema, ModelType, parakeetLoadConfigSchema, LEGACY_PARAKEET_ONNX_MODEL_CONFIG_FIELDS, ADDON_ASR } from '../../../schemas/index.js';
3
- import { ModelLoadFailedError, TranscriptionFailedError, LegacyParakeetModelDeprecatedError } from '../../../errors/index.js';
3
+ import { ModelLoadFailedError, LegacyParakeetModelDeprecatedError } from '../../../errors/index.js';
4
4
  import { transcribe, transcribeStream } from '../../ops/transcribe.js';
5
5
  import { attachModelExecutionMs } from '../../../profiling/model-execution.js';
6
+ import { attachBackendDiagnostics } from '../../../profiling/backend-diagnostics.js';
6
7
  import { buildParakeetEngineConfig } from '../asr-ggml/config.js';
7
8
  import { createAsrModelLogger } from '../asr-ggml/logging.js';
8
9
  function resolveParakeetConfig(cfg) {
@@ -33,6 +34,7 @@ export const parakeetPlugin = definePlugin({
33
34
  displayName: 'Parakeet (NVIDIA NeMo GGML)',
34
35
  addonPackage: ADDON_ASR,
35
36
  loadConfigSchema: parakeetLoadConfigSchema,
37
+ assessFit: ASRGgml.assessFit,
36
38
  resolveConfig(cfg) {
37
39
  return resolveParakeetConfig(cfg);
38
40
  },
@@ -46,30 +48,48 @@ export const parakeetPlugin = definePlugin({
46
48
  streaming: true,
47
49
  cancel: { scope: 'model', hard: true },
48
50
  handler: async function* (request) {
49
- if (request.metadata === true) {
50
- throw new TranscriptionFailedError(`Parakeet transcription does not support metadata: true; only the whisper engine emits per-segment metadata. Use a whisper model to receive segments.`);
51
- }
52
- const stream = transcribe({
53
- modelId: request.modelId,
54
- audioChunk: request.audioChunk,
55
- prompt: request.prompt
56
- }, request.requestId);
51
+ // The native serializer emits start/end/id/toAppend/isEndOfTurn/
52
+ // startsWord on every segment, so metadata mode is forwarded exactly
53
+ // as the whisper plugin does. Branching the call keeps the op's
54
+ // `metadata: true` overload, which is what types results as segments.
55
+ const metadata = request.metadata === true;
56
+ const stream = metadata
57
+ ? transcribe({
58
+ modelId: request.modelId,
59
+ audioChunk: request.audioChunk,
60
+ prompt: request.prompt,
61
+ metadata: true
62
+ }, request.requestId)
63
+ : transcribe({
64
+ modelId: request.modelId,
65
+ audioChunk: request.audioChunk,
66
+ prompt: request.prompt
67
+ }, request.requestId);
57
68
  try {
58
69
  let result = await stream.next();
59
70
  while (!result.done) {
60
- yield {
61
- type: 'transcribe',
62
- text: result.value
63
- };
71
+ yield metadata
72
+ ? {
73
+ type: 'transcribe',
74
+ segment: result.value
75
+ }
76
+ : {
77
+ type: 'transcribe',
78
+ text: result.value
79
+ };
64
80
  result = await stream.next();
65
81
  }
66
- const { modelExecutionMs, stats } = result.value;
67
- yield attachModelExecutionMs({
82
+ const { modelExecutionMs, stats, diagnostics } = result.value;
83
+ // The field is what reaches an RPC client; the symbol is what the
84
+ // profiling layer reads to set `event.backend`, as audiogen does.
85
+ const terminal = attachModelExecutionMs({
68
86
  type: 'transcribe',
69
87
  text: '',
70
88
  done: true,
71
- ...(stats && { stats })
89
+ ...(stats && { stats }),
90
+ ...(diagnostics && { diagnostics })
72
91
  }, modelExecutionMs);
92
+ yield diagnostics ? attachBackendDiagnostics(terminal, diagnostics) : terminal;
73
93
  }
74
94
  finally {
75
95
  await stream.return?.(undefined);
@@ -90,15 +110,15 @@ export const parakeetPlugin = definePlugin({
90
110
  // the duplex RPC writer — backpressure is not yet characterised. Pair
91
111
  // the fix with request-lifecycle `cancel({ requestId })` routing.
92
112
  handler: async function* (request, inputStream) {
93
- if (request.metadata === true) {
94
- throw new TranscriptionFailedError(`Parakeet transcribeStream does not support metadata: true; only the whisper engine emits per-segment metadata.`);
95
- }
113
+ const metadata = request.metadata === true;
96
114
  const streamOpts = {
97
115
  ...(request.parakeetStreamingConfig && {
98
116
  parakeetStreamingConfig: request.parakeetStreamingConfig
99
117
  })
100
118
  };
101
- const iterator = transcribeStream(request.modelId, inputStream, undefined, false, streamOpts, request.requestId);
119
+ const iterator = metadata
120
+ ? transcribeStream(request.modelId, inputStream, undefined, true, streamOpts, request.requestId)
121
+ : transcribeStream(request.modelId, inputStream, undefined, false, streamOpts, request.requestId);
102
122
  try {
103
123
  let result = await iterator.next();
104
124
  while (!result.done) {
@@ -113,19 +133,28 @@ export const parakeetPlugin = definePlugin({
113
133
  result = await iterator.next();
114
134
  continue;
115
135
  }
116
- yield {
117
- type: 'transcribeStream',
118
- text: value
119
- };
136
+ yield metadata
137
+ ? {
138
+ type: 'transcribeStream',
139
+ segment: value
140
+ }
141
+ : {
142
+ type: 'transcribeStream',
143
+ text: value
144
+ };
120
145
  result = await iterator.next();
121
146
  }
122
- const { modelExecutionMs, stats } = result.value;
123
- yield attachModelExecutionMs({
147
+ const { modelExecutionMs, stats, diagnostics } = result.value;
148
+ // The field is what reaches an RPC client; the symbol is what the
149
+ // profiling layer reads to set `event.backend`, as audiogen does.
150
+ const terminal = attachModelExecutionMs({
124
151
  type: 'transcribeStream',
125
152
  text: '',
126
153
  done: true,
127
- ...(stats && { stats })
154
+ ...(stats && { stats }),
155
+ ...(diagnostics && { diagnostics })
128
156
  }, modelExecutionMs);
157
+ yield diagnostics ? attachBackendDiagnostics(terminal, diagnostics) : terminal;
129
158
  }
130
159
  finally {
131
160
  await iterator.return?.(undefined);
@@ -1,3 +1,4 @@
1
+ import { assessFit as diffusionAssessFit } from '@qvac/diffusion-cpp';
1
2
  import { type CreateModelParams, type PluginModelResult, type ResolveContext, type ResolveResult, type SdcppConfig } from '../../../schemas/index.js';
2
3
  type DiffusionArtifactKey = 'clipLModelPath' | 'clipGModelPath' | 'clipVisionModelPath' | 't5XxlModelPath' | 'llmModelPath' | 'vaeModelPath' | 'highNoiseDiffusionModelPath' | 'uncondModelPath' | 'audioVaeModelPath' | 'embeddingsConnectorsModelPath' | 'esrganModelPath' | 'taehvModelPath' | 'seedScenePath';
3
4
  /**
@@ -19,8 +20,8 @@ export declare const diffusionPlugin: {
19
20
  }>>;
20
21
  threads: import("zod").ZodOptional<import("zod").ZodNumber>;
21
22
  device: import("zod").ZodOptional<import("zod").ZodEnum<{
22
- gpu: "gpu";
23
23
  cpu: "cpu";
24
+ gpu: "gpu";
24
25
  }>>;
25
26
  'main-gpu': import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodNumber, import("zod").ZodEnum<{
26
27
  integrated: "integrated";
@@ -554,6 +555,10 @@ export declare const diffusionPlugin: {
554
555
  numFramePerBlock: import("zod").ZodOptional<import("zod").ZodNumber>;
555
556
  localAttnSize: import("zod").ZodOptional<import("zod").ZodNumber>;
556
557
  offloadParamsToCpu: import("zod").ZodOptional<import("zod").ZodBoolean>;
558
+ paramsBackend: import("zod").ZodOptional<import("zod").ZodString>;
559
+ maxVram: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodNumber, import("zod").ZodString]>>;
560
+ streamLayers: import("zod").ZodOptional<import("zod").ZodBoolean>;
561
+ verbosity: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodLiteral<0>, import("zod").ZodLiteral<1>, import("zod").ZodLiteral<2>, import("zod").ZodLiteral<3>]>>;
557
562
  frameJpegQuality: import("zod").ZodOptional<import("zod").ZodNumber>;
558
563
  kvCache: import("zod").ZodOptional<import("zod").ZodBoolean>;
559
564
  profile: import("zod").ZodOptional<import("zod").ZodBoolean>;
@@ -605,6 +610,7 @@ export declare const diffusionPlugin: {
605
610
  threads: import("zod").ZodOptional<import("zod").ZodUnion<readonly [import("zod").ZodLiteral<-1>, import("zod").ZodNumber]>>;
606
611
  }, import("zod/v4/core").$strict>>;
607
612
  }, import("zod/v4/core").$strip>;
613
+ assessFit: typeof diffusionAssessFit;
608
614
  resolveConfig(cfg: SdcppConfig, ctx: ResolveContext): Promise<ResolveResult<SdcppConfig, DiffusionArtifactKey>>;
609
615
  createModel(params: CreateModelParams): PluginModelResult;
610
616
  handlers: {
@@ -846,8 +852,8 @@ export declare const diffusionPlugin: {
846
852
  height: import("zod").ZodOptional<import("zod").ZodNumber>;
847
853
  repeats: import("zod").ZodOptional<import("zod").ZodNumber>;
848
854
  backendDevice: import("zod").ZodOptional<import("zod").ZodEnum<{
849
- gpu: "gpu";
850
855
  cpu: "cpu";
856
+ gpu: "gpu";
851
857
  }>>;
852
858
  }, import("zod/v4/core").$strip>>;
853
859
  }, import("zod/v4/core").$strip>>;
@@ -1,4 +1,4 @@
1
- import ImgStableDiffusion, { EsrganUpscaler, VideoStableDiffusion } from '@qvac/diffusion-cpp';
1
+ import ImgStableDiffusion, { EsrganUpscaler, VideoStableDiffusion, assessFit as diffusionAssessFit } from '@qvac/diffusion-cpp';
2
2
  import { definePlugin, defineHandler, sdcppConfigSchema, diffusionRequestSchema, diffusionStreamResponseSchema, videoRequestSchema, videoStreamResponseSchema, upscaleRequestSchema, upscaleStreamResponseSchema, worldSceneRequestSchema, worldSceneStreamResponseSchema, worldStepRequestSchema, worldStepStreamResponseSchema, ModelType, ADDON_DIFFUSION } from '../../../schemas/index.js';
3
3
  import { createStreamLogger, registerAddonLogger, getEngineLogger } from '../../../logging/index.js';
4
4
  import { ModelLoadFailedError } from '../../../errors/index.js';
@@ -51,6 +51,7 @@ export const diffusionPlugin = definePlugin({
51
51
  displayName: 'Image Generation & Upscaling (stable-diffusion.cpp)',
52
52
  addonPackage: ADDON_DIFFUSION,
53
53
  loadConfigSchema: sdcppConfigSchema,
54
+ assessFit: diffusionAssessFit,
54
55
  async resolveConfig(cfg, ctx) {
55
56
  if (cfg.uncondModelSrc && cfg.mode !== 'diffusion') {
56
57
  throw new ModelLoadFailedError('modelConfig.uncondModelSrc is Ideogram 4 diffusion only. ' +
@@ -112,7 +113,11 @@ export const diffusionPlugin = definePlugin({
112
113
  'main-gpu': 'world.backend',
113
114
  backend: 'world.backend',
114
115
  threads: 'world.threads',
115
- offload_to_cpu: 'world.offloadParamsToCpu'
116
+ offload_to_cpu: 'world.offloadParamsToCpu',
117
+ params_backend: 'world.paramsBackend',
118
+ max_vram: 'world.maxVram',
119
+ stream_layers: 'world.streamLayers',
120
+ verbosity: 'world.verbosity'
116
121
  };
117
122
  const unsupported = Object.keys(cfg).filter((key) => !worldSupported.has(key) && cfg[key] !== undefined);
118
123
  if (unsupported.length > 0) {
@@ -1,6 +1,8 @@
1
1
  import type { TextToSpeechStreamRequest, TtsRequest } from '../../../../schemas/index.js';
2
2
  export type ParlerJobOptions = Pick<TtsRequest, 'description' | 'voiceDescription' | 'voice' | 'emotion' | 'pitch' | 'pace' | 'expressivity' | 'noise' | 'reverb' | 'quality'>;
3
3
  type ParlerJobOptionsRequest = TtsRequest | TextToSpeechStreamRequest;
4
+ /** The loaded addon's engine (`TTSGgml#getEngineType()`), when it reports one. */
5
+ export declare function getTtsEngineType(model: unknown): string | undefined;
4
6
  export declare function getParlerJobOptions(request: ParlerJobOptionsRequest): ParlerJobOptions;
5
7
  export declare function assertParlerJobOptionsSupported(model: unknown, options: ParlerJobOptions, handlerName: 'textToSpeech' | 'textToSpeechStream'): void;
6
8
  export {};
@@ -6,6 +6,10 @@ function hasEngineType(model) {
6
6
  'getEngineType' in model &&
7
7
  typeof model.getEngineType === 'function');
8
8
  }
9
+ /** The loaded addon's engine (`TTSGgml#getEngineType()`), when it reports one. */
10
+ export function getTtsEngineType(model) {
11
+ return hasEngineType(model) ? model.getEngineType() : undefined;
12
+ }
9
13
  export function getParlerJobOptions(request) {
10
14
  return {
11
15
  ...(request.description !== undefined ? { description: request.description } : {}),
@@ -30,7 +34,7 @@ export function assertParlerJobOptionsSupported(model, options, handlerName) {
30
34
  const setKeys = Object.keys(options);
31
35
  if (setKeys.length === 0)
32
36
  return;
33
- const engine = hasEngineType(model) ? model.getEngineType() : undefined;
37
+ const engine = getTtsEngineType(model);
34
38
  if (engine === 'parler')
35
39
  return;
36
40
  if (engine === 'cosyvoice3') {
@@ -54,9 +54,9 @@ import { getModel } from '../../../../runtime/model-registry.js';
54
54
  import { ttsRequestSchema } from '../../../../schemas/index.js';
55
55
  import { nowMs } from '../../../../profiling/index.js';
56
56
  import { buildStreamResult, hasDefinedValues } from '../../../../profiling/model-execution.js';
57
- import { TextToSpeechFailedError } from '../../../../errors/index.js';
58
- import { collectTtsStats, chunkMetadata } from '../../../../utils/tts-stats.js';
59
- import { assertParlerJobOptionsSupported, getParlerJobOptions } from './parler-options.js';
57
+ import { PluginRequestValidationFailedError, TextToSpeechFailedError } from '../../../../errors/index.js';
58
+ import { appendPcm, collectTtsStats, chunkMetadata } from '../../../../utils/tts-stats.js';
59
+ import { assertParlerJobOptionsSupported, getParlerJobOptions, getTtsEngineType } from './parler-options.js';
60
60
  import { bindTtsCancel, cancelIfAborted } from './cancel-binding.js';
61
61
  function hasRunStream(model) {
62
62
  return (typeof model === 'object' &&
@@ -86,6 +86,20 @@ export async function* textToSpeech(params) {
86
86
  const parlerJobOptions = getParlerJobOptions(request);
87
87
  const model = getModel(modelId);
88
88
  assertParlerJobOptionsSupported(model, parlerJobOptions, 'textToSpeech');
89
+ // MOSS streams natively: one job over the whole text, emitting a chunk every
90
+ // `streamChunkTokens` codec frames while the backbone is still generating.
91
+ // The sentence chunker behind `streamOutput` would restart the prompt per
92
+ // sentence, which a MOSS-TTSD dialogue cannot do (every job has to open
93
+ // with the reference transcripts, so the addon rejects it) and which would
94
+ // apply `durationTokens` to each sentence rather than to the text.
95
+ // `sentenceStream: true` still asks for the chunker explicitly.
96
+ const nativeStream = stream && !sentenceStream && getTtsEngineType(model) === 'moss';
97
+ // The chunker is what reads these, so on the native path they would be
98
+ // dropped without a trace.
99
+ if (nativeStream &&
100
+ (sentenceStreamLocale !== undefined || sentenceStreamMaxChunkScalars !== undefined)) {
101
+ throw new PluginRequestValidationFailedError('textToSpeech', 'sentenceStreamLocale and sentenceStreamMaxChunkScalars need the sentence chunker; MOSS streams the whole text natively, so they only apply with sentenceStream: true');
102
+ }
89
103
  const ctx = __addDisposableResource(env_1, await bindTtsCancel(model, modelId, request.requestId), true);
90
104
  if (ctx.signal.aborted)
91
105
  return { ...buildStreamResult(0), cancelled: true };
@@ -107,7 +121,7 @@ export async function* textToSpeech(params) {
107
121
  : undefined;
108
122
  const response = await model.runStream(text, streamOpts);
109
123
  if (!stream) {
110
- let completeBuffer = [];
124
+ const completeBuffer = [];
111
125
  let sampleRate;
112
126
  try {
113
127
  for await (const data of response.iterate()) {
@@ -116,7 +130,7 @@ export async function* textToSpeech(params) {
116
130
  // lunte-disable-next-line eqeqeq -- `!= null` intentionally matches null and undefined
117
131
  if (data.outputArray != null) {
118
132
  sampleRate ??= data.sampleRate;
119
- completeBuffer = completeBuffer.concat(Array.from(data.outputArray));
133
+ appendPcm(completeBuffer, data.outputArray);
120
134
  }
121
135
  }
122
136
  }
@@ -147,6 +161,7 @@ export async function* textToSpeech(params) {
147
161
  }
148
162
  return finish(modelStart, response, ctx.signal.aborted);
149
163
  }
164
+ const sentenceChunked = stream && !nativeStream;
150
165
  const response = (await model.run({
151
166
  input: text,
152
167
  // The addon's job field is `type`, and its native layer accepts only
@@ -154,25 +169,27 @@ export async function* textToSpeech(params) {
154
169
  // it. Pin it here too, so the request's `inputType` (which the schema
155
170
  // does not constrain) behaves the same on every path.
156
171
  type: 'text',
157
- ...(stream ? { streamOutput: true } : {}),
172
+ ...(sentenceChunked ? { streamOutput: true } : {}),
158
173
  // `run({ streamOutput: true })` runs the same chunker as `runStream()`, so
159
174
  // the chunking knobs apply here too — they used to be honoured only on the
160
175
  // sentenceStream path.
161
- ...(stream && sentenceStreamLocale !== undefined ? { locale: sentenceStreamLocale } : {}),
162
- ...(stream && sentenceStreamMaxChunkScalars !== undefined
176
+ ...(sentenceChunked && sentenceStreamLocale !== undefined
177
+ ? { locale: sentenceStreamLocale }
178
+ : {}),
179
+ ...(sentenceChunked && sentenceStreamMaxChunkScalars !== undefined
163
180
  ? { maxChunkScalars: sentenceStreamMaxChunkScalars }
164
181
  : {}),
165
182
  ...parlerJobOptions
166
183
  }));
167
184
  if (!stream) {
168
- let completeBuffer = [];
185
+ const completeBuffer = [];
169
186
  let sampleRate;
170
187
  try {
171
188
  for await (const data of response.iterate()) {
172
189
  if (await cancelIfAborted(model, ctx.signal))
173
190
  continue;
174
191
  sampleRate ??= data.sampleRate;
175
- completeBuffer = completeBuffer.concat(Array.from(data.outputArray));
192
+ appendPcm(completeBuffer, data.outputArray);
176
193
  }
177
194
  }
178
195
  catch (error) {