@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
@@ -427,6 +427,26 @@ export const MODEL_RESOURCE_PROFILES = {
427
427
  engine: 'nmtcpp-translation',
428
428
  artifactBytes: 2248
429
429
  },
430
+ '166226feb6144f6e416e4936706eb5f9183526a05141410156c578c48b4bd4f1': {
431
+ schemaVersion: 1,
432
+ engine: 'llamacpp-completion',
433
+ artifactBytes: 3066385344,
434
+ ggufFacts: {
435
+ architecture: 'qwen35',
436
+ blockCount: 32,
437
+ headCount: 16,
438
+ headCountKv: 4,
439
+ keyLength: 256,
440
+ valueLength: 256,
441
+ embeddingLength: 2560,
442
+ contextLength: 262144,
443
+ fullAttentionInterval: 4,
444
+ ssmStateSize: 128,
445
+ ssmConvKernel: 4,
446
+ ssmInnerSize: 4096,
447
+ ssmGroupCount: 16
448
+ }
449
+ },
430
450
  '1686f94838ace314719e945606a09a66dafd40847dd3fd352ad3f7a969aac62f': {
431
451
  schemaVersion: 1,
432
452
  engine: 'nmtcpp-translation',
@@ -745,6 +765,26 @@ export const MODEL_RESOURCE_PROFILES = {
745
765
  engine: 'nmtcpp-translation',
746
766
  artifactBytes: 653915115
747
767
  },
768
+ '296d6426b922a09c6615a81dee6e0aaae6f1f608fd2c320316f1e6ce09314aac': {
769
+ schemaVersion: 1,
770
+ engine: 'llamacpp-completion',
771
+ artifactBytes: 5157833664,
772
+ ggufFacts: {
773
+ architecture: 'qwen35',
774
+ blockCount: 32,
775
+ headCount: 16,
776
+ headCountKv: 4,
777
+ keyLength: 256,
778
+ valueLength: 256,
779
+ embeddingLength: 2560,
780
+ contextLength: 262144,
781
+ fullAttentionInterval: 4,
782
+ ssmStateSize: 128,
783
+ ssmConvKernel: 4,
784
+ ssmInnerSize: 4096,
785
+ ssmGroupCount: 16
786
+ }
787
+ },
748
788
  '29940d98d42b91fbd05ce489f3ecf7c72f0a42f027e4875919a28fb4c04ea2cf': {
749
789
  schemaVersion: 1,
750
790
  engine: 'whispercpp-transcription',
@@ -1425,6 +1465,26 @@ export const MODEL_RESOURCE_PROFILES = {
1425
1465
  artifactBytes: 37014606,
1426
1466
  assumptions: ['artifactBytes sums 4 companion-set files']
1427
1467
  },
1468
+ '4af8ee1df3ec9008f763ebe95e6f21df3acd8d42c541feeb13314ca22e560afc': {
1469
+ schemaVersion: 1,
1470
+ engine: 'llamacpp-completion',
1471
+ artifactBytes: 672329792,
1472
+ ggufFacts: {
1473
+ architecture: 'qwen35',
1474
+ blockCount: 24,
1475
+ headCount: 8,
1476
+ headCountKv: 2,
1477
+ keyLength: 256,
1478
+ valueLength: 256,
1479
+ embeddingLength: 1024,
1480
+ contextLength: 262144,
1481
+ fullAttentionInterval: 4,
1482
+ ssmStateSize: 128,
1483
+ ssmConvKernel: 4,
1484
+ ssmInnerSize: 2048,
1485
+ ssmGroupCount: 16
1486
+ }
1487
+ },
1428
1488
  '4b0ce319e7c8f7e1836ce18cb6522c83268e8cff54f2a7f5dcef270250a44475': {
1429
1489
  schemaVersion: 1,
1430
1490
  engine: 'nmtcpp-translation',
@@ -2804,6 +2864,26 @@ export const MODEL_RESOURCE_PROFILES = {
2804
2864
  engine: 'nmtcpp-translation',
2805
2865
  artifactBytes: 3075948
2806
2866
  },
2867
+ '940436f99196c1183d0a38addbbb2c643cb5edfcb939003dacc8c4bd68079f85': {
2868
+ schemaVersion: 1,
2869
+ engine: 'llamacpp-completion',
2870
+ artifactBytes: 2552356928,
2871
+ ggufFacts: {
2872
+ architecture: 'qwen35',
2873
+ blockCount: 24,
2874
+ headCount: 8,
2875
+ headCountKv: 2,
2876
+ keyLength: 256,
2877
+ valueLength: 256,
2878
+ embeddingLength: 2048,
2879
+ contextLength: 262144,
2880
+ fullAttentionInterval: 4,
2881
+ ssmStateSize: 128,
2882
+ ssmConvKernel: 4,
2883
+ ssmInnerSize: 2048,
2884
+ ssmGroupCount: 16
2885
+ }
2886
+ },
2807
2887
  '94b660f37c9c29a904b48498a7202846faaeef16f30668f356336c29cefc0f63': {
2808
2888
  schemaVersion: 1,
2809
2889
  engine: 'nmtcpp-translation',
@@ -3263,6 +3343,11 @@ export const MODEL_RESOURCE_PROFILES = {
3263
3343
  engine: 'nmtcpp-translation',
3264
3344
  artifactBytes: 4197836
3265
3345
  },
3346
+ af090772797632dcea0546985b112f86a9a10e0d87d10a2b74bd6ae57464aded: {
3347
+ schemaVersion: 1,
3348
+ engine: 'parakeet-transcription',
3349
+ artifactBytes: 1429807168
3350
+ },
3266
3351
  af151db26e3b0a87e805d899d3f105aa0a11f9124893d47ca3f39e47b27cfa8f: {
3267
3352
  schemaVersion: 1,
3268
3353
  engine: 'sdcpp-generation',
@@ -3436,6 +3521,26 @@ export const MODEL_RESOURCE_PROFILES = {
3436
3521
  engine: 'parakeet-transcription',
3437
3522
  artifactBytes: 1371674432
3438
3523
  },
3524
+ b5056635be0c9bb832f0c964dc063bd4b4462e253f9f9d94eebf0aa0960b7798: {
3525
+ schemaVersion: 1,
3526
+ engine: 'llamacpp-completion',
3527
+ artifactBytes: 1082015808,
3528
+ ggufFacts: {
3529
+ architecture: 'qwen35',
3530
+ blockCount: 24,
3531
+ headCount: 8,
3532
+ headCountKv: 2,
3533
+ keyLength: 256,
3534
+ valueLength: 256,
3535
+ embeddingLength: 1024,
3536
+ contextLength: 262144,
3537
+ fullAttentionInterval: 4,
3538
+ ssmStateSize: 128,
3539
+ ssmConvKernel: 4,
3540
+ ssmInnerSize: 2048,
3541
+ ssmGroupCount: 16
3542
+ }
3543
+ },
3439
3544
  b5310340b3a23d31655d7119d100d5df1b2d8ee17b3ca8b0a23ad7e9eb5fa705: {
3440
3545
  schemaVersion: 1,
3441
3546
  engine: 'llamacpp-completion',
@@ -4664,6 +4769,11 @@ export const MODEL_RESOURCE_PROFILES = {
4664
4769
  engine: 'nmtcpp-translation',
4665
4770
  artifactBytes: 4724124
4666
4771
  },
4772
+ f6ad1b9912c03c95b3377eaf0fc2fee635a4e4904409d8efac1944343d9f3360: {
4773
+ schemaVersion: 1,
4774
+ engine: 'parakeet-transcription',
4775
+ artifactBytes: 761593408
4776
+ },
4667
4777
  f6b6c6d5cfa6f00d964eeb7add28eb14ce7481734d506b90681007678cd2c484: {
4668
4778
  schemaVersion: 1,
4669
4779
  engine: 'llamacpp-completion',
@@ -4783,6 +4893,26 @@ export const MODEL_RESOURCE_PROFILES = {
4783
4893
  engine: 'whispercpp-transcription',
4784
4894
  artifactBytes: 77691730
4785
4895
  },
4896
+ fc2a80532e58e0655efd9e68c5b7cb2c4373222e736bcc88de648502a4e30361: {
4897
+ schemaVersion: 1,
4898
+ engine: 'llamacpp-completion',
4899
+ artifactBytes: 1560461376,
4900
+ ggufFacts: {
4901
+ architecture: 'qwen35',
4902
+ blockCount: 24,
4903
+ headCount: 8,
4904
+ headCountKv: 2,
4905
+ keyLength: 256,
4906
+ valueLength: 256,
4907
+ embeddingLength: 2048,
4908
+ contextLength: 262144,
4909
+ fullAttentionInterval: 4,
4910
+ ssmStateSize: 128,
4911
+ ssmConvKernel: 4,
4912
+ ssmInnerSize: 2048,
4913
+ ssmGroupCount: 16
4914
+ }
4915
+ },
4786
4916
  fc70a6c6eed7d2f82ed48cbd52cc7118b249eff8b91112ef9cbfca6813a1eefa: {
4787
4917
  schemaVersion: 1,
4788
4918
  engine: 'llamacpp-completion',
@@ -4858,6 +4988,11 @@ export const MODEL_RESOURCE_PROFILES = {
4858
4988
  ssmGroupCount: 16
4859
4989
  }
4860
4990
  },
4991
+ fee699b3257ce0cfd249f490ab9f9791d9e0f6f8781e4354db2aa5584e594c82: {
4992
+ schemaVersion: 1,
4993
+ engine: 'parakeet-transcription',
4994
+ artifactBytes: 405212736
4995
+ },
4861
4996
  fefd91bc5551877762cd0d0cbe583080f8e636d5aa65a9a5f42b369654df778b: {
4862
4997
  schemaVersion: 1,
4863
4998
  engine: 'llamacpp-completion',
@@ -0,0 +1,2 @@
1
+ import { type AudioEditStreamRequest, type AudioEditStreamResponse } from '../../../../schemas/audio-gen.js';
2
+ export declare function audioEditStream(request: AudioEditStreamRequest): AsyncGenerator<AudioEditStreamResponse>;
@@ -0,0 +1,65 @@
1
+ import { RepaintMode } from '@qvac/audiogen-ggml';
2
+ import { AUDIOGEN_INPUT_CHANNELS, AUDIOGEN_INPUT_SAMPLE_RATE } from '../../../../schemas/audio-gen.js';
3
+ import { resolveAudioGenPcm } from './audio-gen-input.js';
4
+ import { streamAudioGenRun } from './audio-gen-run.js';
5
+ // The wire vocabulary is the addon's own enum values; the map keeps the
6
+ // string-enum typing honest instead of casting across the boundary.
7
+ const REPAINT_MODES = {
8
+ conservative: RepaintMode.Conservative,
9
+ balanced: RepaintMode.Balanced,
10
+ aggressive: RepaintMode.Aggressive
11
+ };
12
+ export function audioEditStream(request) {
13
+ return streamAudioGenRun({
14
+ type: 'audioEditStream',
15
+ request,
16
+ async start(model, ctx) {
17
+ // The source is decoded before the run is admitted so the model slot is
18
+ // never held by a request that fails on input decoding. Editing is the
19
+ // one path where the addon demands normalized samples, so the resolver
20
+ // settles `[-1, 1]` here for raw and decoded input alike.
21
+ const pcm = await resolveAudioGenPcm(request.sourceAudio, 'sourceAudio', {
22
+ requireNormalized: true
23
+ });
24
+ if (ctx.signal.aborted)
25
+ return undefined;
26
+ // The addon validates each operation as it is chained (Flow-Edit on a
27
+ // turbo DiT, repaint ranges inside the source, ...); those throws
28
+ // surface as the request's error before any native work starts.
29
+ const session = model.edit({
30
+ pcm,
31
+ sampleRate: AUDIOGEN_INPUT_SAMPLE_RATE,
32
+ channels: AUDIOGEN_INPUT_CHANNELS
33
+ });
34
+ for (const operation of request.operations)
35
+ appendOperation(session, operation);
36
+ return session.run({ ...(request.seed !== undefined && { seed: request.seed }) });
37
+ }
38
+ });
39
+ }
40
+ function appendOperation(session, operation) {
41
+ if (operation.type === 'flow-edit') {
42
+ session.flowEdit({
43
+ from: {
44
+ caption: operation.from.caption,
45
+ ...(operation.from.lyrics !== undefined && { lyrics: operation.from.lyrics })
46
+ },
47
+ to: {
48
+ caption: operation.to.caption,
49
+ ...(operation.to.lyrics !== undefined && { lyrics: operation.to.lyrics })
50
+ },
51
+ ...(operation.nMin !== undefined && { nMin: operation.nMin }),
52
+ ...(operation.nMax !== undefined && { nMax: operation.nMax }),
53
+ ...(operation.nAvg !== undefined && { nAvg: operation.nAvg })
54
+ });
55
+ return;
56
+ }
57
+ session.repaint({
58
+ caption: operation.caption,
59
+ ...(operation.lyrics !== undefined && { lyrics: operation.lyrics }),
60
+ start: operation.start,
61
+ ...(operation.end !== undefined && { end: operation.end }),
62
+ ...(operation.mode !== undefined && { mode: REPAINT_MODES[operation.mode] }),
63
+ ...(operation.strength !== undefined && { strength: operation.strength })
64
+ });
65
+ }
@@ -15,5 +15,13 @@ export type AudioGenAudioInputName = 'referenceAudio' | 'sourceAudio';
15
15
  * Every path is bounded by `AUDIOGEN_INPUT_MAX_SECONDS` and rejects
16
16
  * non-finite samples, so malformed or oversized input fails with
17
17
  * `InvalidAudioInputError` before the engine is invoked.
18
+ *
19
+ * `requireNormalized` additionally holds samples to `[-1, 1]`. Only the edit
20
+ * path needs it — the addon vets an edit source with `requireNormalizedPcm`
21
+ * and everything else with `requireFinitePcm`, which takes any finite sample.
22
+ * It applies to decoded files as much as to raw PCM: decoding does not
23
+ * normalize, so a loud source comes back above 1.0 either way.
18
24
  */
19
- export declare function resolveAudioGenPcm(input: AudioGenAudioInput, name: AudioGenAudioInputName): Promise<Float32Array>;
25
+ export declare function resolveAudioGenPcm(input: AudioGenAudioInput, name: AudioGenAudioInputName, { requireNormalized }?: {
26
+ requireNormalized?: boolean;
27
+ }): Promise<Float32Array>;
@@ -23,18 +23,24 @@ const MAX_DECODED_MONO_BYTES = AUDIOGEN_INPUT_MAX_SECONDS * AUDIOGEN_INPUT_SAMPL
23
23
  * Every path is bounded by `AUDIOGEN_INPUT_MAX_SECONDS` and rejects
24
24
  * non-finite samples, so malformed or oversized input fails with
25
25
  * `InvalidAudioInputError` before the engine is invoked.
26
+ *
27
+ * `requireNormalized` additionally holds samples to `[-1, 1]`. Only the edit
28
+ * path needs it — the addon vets an edit source with `requireNormalizedPcm`
29
+ * and everything else with `requireFinitePcm`, which takes any finite sample.
30
+ * It applies to decoded files as much as to raw PCM: decoding does not
31
+ * normalize, so a loud source comes back above 1.0 either way.
26
32
  */
27
- export async function resolveAudioGenPcm(input, name) {
33
+ export async function resolveAudioGenPcm(input, name, { requireNormalized = false } = {}) {
28
34
  switch (input.type) {
29
35
  case 'base64':
30
- return toStereoFloat32(Buffer.from(input.value, 'base64'), name);
36
+ return toStereoFloat32(Buffer.from(input.value, 'base64'), name, requireNormalized);
31
37
  case 'filePath': {
32
38
  const filePath = input.value;
33
39
  if (!needsDecoding(filePath)) {
34
- return toStereoFloat32(await readRawPcmFile(filePath, name), name);
40
+ return toStereoFloat32(await readRawPcmFile(filePath, name), name, requireNormalized);
35
41
  }
36
42
  await assertDecodableFileSize(filePath, name);
37
- return monoToStereo(await decodeMonoFloat32(filePath, name), name);
43
+ return monoToStereo(await decodeMonoFloat32(filePath, name), name, requireNormalized);
38
44
  }
39
45
  default:
40
46
  throw new InvalidAudioInputError(`${name} must be a file path or raw PCM bytes`);
@@ -103,7 +109,7 @@ async function decodeMonoFloat32(filePath, name) {
103
109
  }
104
110
  return asFloat32(chunks.length === 1 ? chunks[0] : Buffer.concat(chunks, total));
105
111
  }
106
- function monoToStereo(mono, name) {
112
+ function monoToStereo(mono, name, requireNormalized) {
107
113
  if (mono.length === 0) {
108
114
  throw new InvalidAudioInputError(`${name} decoded to no usable audio`);
109
115
  }
@@ -112,21 +118,29 @@ function monoToStereo(mono, name) {
112
118
  const sample = mono[frame];
113
119
  if (!Number.isFinite(sample))
114
120
  throw nonFiniteError(name);
121
+ if (requireNormalized && (sample < -1 || sample > 1))
122
+ throw outOfRangeError(name, sample);
115
123
  stereo[frame * AUDIOGEN_INPUT_CHANNELS] = sample;
116
124
  stereo[frame * AUDIOGEN_INPUT_CHANNELS + 1] = sample;
117
125
  }
118
126
  return stereo;
119
127
  }
120
- function toStereoFloat32(bytes, name) {
128
+ function toStereoFloat32(bytes, name, requireNormalized) {
121
129
  if (bytes.byteLength === 0 || bytes.byteLength % STEREO_FRAME_BYTES !== 0) {
122
130
  throw new InvalidAudioInputError(`${name} must be non-empty interleaved stereo ${AUDIOGEN_INPUT_SAMPLE_RATE} Hz Float32 PCM ` +
123
131
  `(byte length a multiple of ${STEREO_FRAME_BYTES}, got ${bytes.byteLength})`);
124
132
  }
125
133
  assertWithinLimit(bytes.byteLength, MAX_STEREO_BYTES, STEREO_FRAME_BYTES, name);
126
134
  const pcm = asFloat32(bytes);
135
+ // One pass settles every sample-level invariant this input needs: at the
136
+ // documented 600 s cap a second walk would be tens of millions of extra
137
+ // comparisons.
127
138
  for (let index = 0; index < pcm.length; index++) {
128
- if (!Number.isFinite(pcm[index]))
139
+ const sample = pcm[index];
140
+ if (!Number.isFinite(sample))
129
141
  throw nonFiniteError(name);
142
+ if (requireNormalized && (sample < -1 || sample > 1))
143
+ throw outOfRangeError(name, sample);
130
144
  }
131
145
  return pcm;
132
146
  }
@@ -151,3 +165,6 @@ function assertWithinLimit(bytes, limit, frameBytes, name, filePath) {
151
165
  function nonFiniteError(name) {
152
166
  return new InvalidAudioInputError(`${name} must contain only finite samples`);
153
167
  }
168
+ function outOfRangeError(name, sample) {
169
+ return new InvalidAudioInputError(`${name} must contain samples in [-1, 1] (got ${sample})`);
170
+ }
@@ -0,0 +1,46 @@
1
+ import { AudioGen } from '@qvac/audiogen-ggml';
2
+ import { type AudioGenStreamResponse, type AudioUnderstandResponse } from '../../../../schemas/audio-gen.js';
3
+ import { type Logger } from '../../../../logging/index.js';
4
+ import { type RequestContext } from '../../../../runtime/index.js';
5
+ /**
6
+ * Wire types that stream an AudioGen run: generation, source-driven editing,
7
+ * and the reverse pipeline behind `audioUnderstand()`.
8
+ */
9
+ export type AudioGenRunType = 'audioGenStream' | 'audioEditStream' | 'audioUnderstand';
10
+ /**
11
+ * A stream frame for `TType`. Generation and editing share every field but
12
+ * `type`; understanding carries a description instead of PCM.
13
+ *
14
+ * The frames below are built as object literals inside a generic function, so
15
+ * TypeScript cannot match them to the narrowed conditional type on its own —
16
+ * each `yield` asserts the frame it just built. The shapes are pinned at the
17
+ * boundary regardless: every frame is parsed against this handler's response
18
+ * schema before it reaches a caller.
19
+ */
20
+ export type AudioGenRunFrame<TType extends AudioGenRunType> = TType extends 'audioUnderstand' ? AudioUnderstandResponse : Omit<AudioGenStreamResponse, 'type'> & {
21
+ type: TType;
22
+ };
23
+ export type AudioGenRunResponse = Awaited<ReturnType<AudioGen['run']>>;
24
+ export interface AudioGenRunOptions<TType extends AudioGenRunType> {
25
+ type: TType;
26
+ request: {
27
+ modelId: string;
28
+ requestId?: string | undefined;
29
+ };
30
+ /**
31
+ * Prepares the inputs and admits the native job once the request holds the
32
+ * model slot. Preparation (decoding audio) can take real time, so the
33
+ * callback must re-check `ctx.signal` after it and return `undefined`
34
+ * without starting the job when an abort fired mid-way: that abort found no
35
+ * active native job to cancel, and starting the run afterwards would leave
36
+ * a ghost generation on the slot.
37
+ */
38
+ start: (model: AudioGen, ctx: RequestContext, logger: Logger) => Promise<AudioGenRunResponse | undefined>;
39
+ }
40
+ /**
41
+ * Runs one AudioGen job under the request registry and streams it as
42
+ * `TType` frames: progress ticks, PCM chunks, then the terminal frame with
43
+ * stats and backend diagnostics. Owns admission (`kind: 'audiogen'`, one
44
+ * active job per model), abort-to-cancel wiring, and the cancelled terminal.
45
+ */
46
+ export declare function streamAudioGenRun<TType extends AudioGenRunType>(options: AudioGenRunOptions<TType>): AsyncGenerator<AudioGenRunFrame<TType>>;
@@ -0,0 +1,218 @@
1
+ var __addDisposableResource = (this && this.__addDisposableResource) || function (env, value, async) {
2
+ if (value !== null && value !== void 0) {
3
+ if (typeof value !== "object" && typeof value !== "function") throw new TypeError("Object expected.");
4
+ var dispose, inner;
5
+ if (async) {
6
+ if (!Symbol.asyncDispose) throw new TypeError("Symbol.asyncDispose is not defined.");
7
+ dispose = value[Symbol.asyncDispose];
8
+ }
9
+ if (dispose === void 0) {
10
+ if (!Symbol.dispose) throw new TypeError("Symbol.dispose is not defined.");
11
+ dispose = value[Symbol.dispose];
12
+ if (async) inner = dispose;
13
+ }
14
+ if (typeof dispose !== "function") throw new TypeError("Object not disposable.");
15
+ if (inner) dispose = function() { try { inner.call(this); } catch (e) { return Promise.reject(e); } };
16
+ env.stack.push({ value: value, dispose: dispose, async: async });
17
+ }
18
+ else if (async) {
19
+ env.stack.push({ async: true });
20
+ }
21
+ return value;
22
+ };
23
+ var __disposeResources = (this && this.__disposeResources) || (function (SuppressedError) {
24
+ return function (env) {
25
+ function fail(e) {
26
+ env.error = env.hasError ? new SuppressedError(e, env.error, "An error was suppressed during disposal.") : e;
27
+ env.hasError = true;
28
+ }
29
+ var r, s = 0;
30
+ function next() {
31
+ while (r = env.stack.pop()) {
32
+ try {
33
+ if (!r.async && s === 1) return s = 0, env.stack.push(r), Promise.resolve().then(next);
34
+ if (r.dispose) {
35
+ var result = r.dispose.call(r.value);
36
+ if (r.async) return s |= 2, Promise.resolve(result).then(next, function(e) { fail(e); return next(); });
37
+ }
38
+ else s |= 1;
39
+ }
40
+ catch (e) {
41
+ fail(e);
42
+ }
43
+ }
44
+ if (s === 1) return env.hasError ? Promise.reject(env.error) : Promise.resolve();
45
+ if (env.hasError) throw env.error;
46
+ }
47
+ return next();
48
+ };
49
+ })(typeof SuppressedError === "function" ? SuppressedError : function (error, suppressed, message) {
50
+ var e = new Error(message);
51
+ return e.name = "SuppressedError", e.error = error, e.suppressed = suppressed, e;
52
+ });
53
+ import { AudioGen, audiogenBackendName, audiogenGpuFallbackReason } from '@qvac/audiogen-ggml';
54
+ import { z } from 'zod';
55
+ import { audioGenStatsSchema } from '../../../../schemas/audio-gen.js';
56
+ import { graphicsDriverSchema } from '../../../../schemas/index.js';
57
+ import { getEngineLogger } from '../../../../logging/index.js';
58
+ import { attachBackendDiagnostics } from '../../../../profiling/backend-diagnostics.js';
59
+ import { getModel } from '../../../../runtime/model-registry.js';
60
+ import { getRequestRegistry, withRequestContext } from '../../../../runtime/index.js';
61
+ import { generateRequestId } from '../../../../runtime/request-id.js';
62
+ import { ModelOperationNotSupportedError } from '../../../../errors/index.js';
63
+ /**
64
+ * Runs one AudioGen job under the request registry and streams it as
65
+ * `TType` frames: progress ticks, PCM chunks, then the terminal frame with
66
+ * stats and backend diagnostics. Owns admission (`kind: 'audiogen'`, one
67
+ * active job per model), abort-to-cancel wiring, and the cancelled terminal.
68
+ */
69
+ export async function* streamAudioGenRun(options) {
70
+ const env_1 = { stack: [], error: void 0, hasError: false };
71
+ try {
72
+ const { type, request } = options;
73
+ const ctx = __addDisposableResource(env_1, await getRequestRegistry().begin({
74
+ requestId: request.requestId ?? generateRequestId(),
75
+ kind: 'audiogen',
76
+ modelId: request.modelId
77
+ }), true);
78
+ const logger = withRequestContext(getEngineLogger(), ctx);
79
+ const candidate = getModel(request.modelId);
80
+ if (!(candidate instanceof AudioGen)) {
81
+ throw new ModelOperationNotSupportedError(request.modelId, 'audiogen-ggml', type, [], []);
82
+ }
83
+ const model = candidate;
84
+ // A queued request can resume from begin() already aborted. It never owned
85
+ // the model slot, so calling the model-scoped cancel here would interrupt
86
+ // the earlier same-model generation that still owns it.
87
+ if (ctx.signal.aborted) {
88
+ yield { type, done: true, stopReason: 'cancelled' };
89
+ return;
90
+ }
91
+ let cancelPromise;
92
+ const onAbort = () => {
93
+ cancelPromise ??= model.cancel().catch((error) => {
94
+ logger.warn(`[cancel] model.cancel() rejected during abort for modelId=${request.modelId}: ${error instanceof Error ? error.message : String(error)}`);
95
+ });
96
+ };
97
+ ctx.signal.addEventListener('abort', onAbort, { once: true });
98
+ if (ctx.signal.aborted)
99
+ onAbort();
100
+ ctx.scope.defer(async () => {
101
+ ctx.signal.removeEventListener('abort', onAbort);
102
+ await cancelPromise;
103
+ });
104
+ let response;
105
+ let streamedUnderstand;
106
+ try {
107
+ response = await options.start(model, ctx, logger);
108
+ if (response !== undefined) {
109
+ for await (const chunk of response.iterate()) {
110
+ if (ctx.signal.aborted)
111
+ break;
112
+ if ('progress' in chunk) {
113
+ yield { type, progress: chunk.progress, done: false };
114
+ continue;
115
+ }
116
+ // `understand()` reports a description instead of audio; generation
117
+ // and editing never produce this item.
118
+ if ('understand' in chunk) {
119
+ // The engine repeats this result on the terminal stats, so the
120
+ // converted form is kept rather than rebuilt from the raw codes.
121
+ streamedUnderstand = toUnderstandResult(chunk.understand);
122
+ yield {
123
+ type,
124
+ understand: streamedUnderstand,
125
+ done: false
126
+ };
127
+ continue;
128
+ }
129
+ if (!('outputArray' in chunk))
130
+ continue;
131
+ const pcm = new Uint8Array(chunk.outputArray.buffer, chunk.outputArray.byteOffset, chunk.outputArray.byteLength);
132
+ yield {
133
+ type,
134
+ data: Buffer.from(pcm).toString('base64'),
135
+ sampleRate: chunk.sampleRate,
136
+ channels: chunk.channels,
137
+ bitsPerSample: Int16Array.BYTES_PER_ELEMENT * 8,
138
+ done: false
139
+ };
140
+ }
141
+ }
142
+ }
143
+ catch (error) {
144
+ if (!ctx.signal.aborted)
145
+ throw error;
146
+ }
147
+ if (ctx.signal.aborted || response === undefined) {
148
+ yield { type, done: true, stopReason: 'cancelled' };
149
+ return;
150
+ }
151
+ const raw = await response.await();
152
+ // `understand.audioCodes` arrives as an Int32Array, which the wire schema —
153
+ // and anything downstream of JSON — needs as a plain array of integers.
154
+ // Both parses below read the normalized object: handing the raw one to
155
+ // either schema fails its `z.array` check on the typed array.
156
+ const normalized = raw.understand
157
+ ? { ...raw, understand: streamedUnderstand ?? toUnderstandResult(raw.understand) }
158
+ : raw;
159
+ const stats = audioGenStatsSchema.parse(normalized);
160
+ const diagnostics = buildBackendDiagnostics(audiogenStats.parse(normalized));
161
+ const terminal = {
162
+ type,
163
+ done: true,
164
+ stopReason: 'completed',
165
+ stats,
166
+ ...(diagnostics && { diagnostics })
167
+ };
168
+ yield diagnostics ? attachBackendDiagnostics(terminal, diagnostics) : terminal;
169
+ }
170
+ catch (e_1) {
171
+ env_1.error = e_1;
172
+ env_1.hasError = true;
173
+ }
174
+ finally {
175
+ const result_1 = __disposeResources(env_1);
176
+ if (result_1)
177
+ await result_1;
178
+ }
179
+ }
180
+ // The addon reports gpuFallbackReason, which the wire shape deliberately does
181
+ // not carry: it reaches callers named, on diagnostics.fallback.reason.
182
+ const audiogenStats = audioGenStatsSchema.extend({
183
+ gpuFallbackReason: z.number().optional()
184
+ });
185
+ /**
186
+ * Normalizes the addon's understand result for the wire: `audioCodes` arrives
187
+ * as an `Int32Array`, which neither the schema nor a JSON transport accepts.
188
+ */
189
+ function toUnderstandResult(result) {
190
+ return { ...result, audioCodes: Array.from(result.audioCodes) };
191
+ }
192
+ /** An unrecognized GPU id yields no diagnostics rather than a guessed backend name. */
193
+ function buildBackendDiagnostics(stats) {
194
+ if (stats.backendDevice === undefined)
195
+ return undefined;
196
+ if (stats.backendDevice !== 1) {
197
+ const fallback = gpuFallbackDetail(stats);
198
+ return { selectedBackend: 'cpu', selectedDevice: 'cpu', ...(fallback && { fallback }) };
199
+ }
200
+ // A 'cpu' name against backendDevice 1 is the addon contradicting itself.
201
+ const selectedBackend = audiogenBackendName(stats.backendId);
202
+ if (selectedBackend === undefined || selectedBackend === 'cpu')
203
+ return undefined;
204
+ const graphicsApi = graphicsDriverSchema.safeParse(selectedBackend);
205
+ return {
206
+ selectedBackend,
207
+ selectedDevice: 'gpu',
208
+ ...(graphicsApi.success && { graphicsApi: graphicsApi.data })
209
+ };
210
+ }
211
+ // `none` and `not-requested` describe a run that never lost a GPU, and an
212
+ // unmapped code must not become a guessed reason.
213
+ function gpuFallbackDetail(stats) {
214
+ const reason = audiogenGpuFallbackReason(stats.gpuFallbackReason);
215
+ if (reason === undefined || reason === 'none' || reason === 'not-requested')
216
+ return undefined;
217
+ return { requestedDevice: 'gpu', reason };
218
+ }
@@ -1,2 +1,2 @@
1
- import { type AudioGenStreamRequest, type AudioGenStreamResponse } from '../../../../schemas/audio-gen.js';
1
+ import type { AudioGenStreamRequest, AudioGenStreamResponse } from '../../../../schemas/audio-gen.js';
2
2
  export declare function audioGenStream(request: AudioGenStreamRequest): AsyncGenerator<AudioGenStreamResponse>;