diff --git a/packages/react-native-executorch/__tests__/extensions/normalize.test.ts b/packages/react-native-executorch/__tests__/extensions/normalize.test.ts new file mode 100644 index 0000000000..bb7cd01d36 --- /dev/null +++ b/packages/react-native-executorch/__tests__/extensions/normalize.test.ts @@ -0,0 +1,18 @@ +import { tensor } from '../../src/core/tensor'; +import { normalize } from '../../src/extensions/cv/ops/image'; + +describe('normalize — options merging', () => { + it('falls back to the defaults for options passed as undefined', () => { + const src = tensor('float32', [1, 1, 2], new Float32Array([255, 51])); + const dst = tensor('float32', [1, 1, 2]); + try { + normalize(src, dst, { alpha: undefined, beta: undefined }); + const [first, second] = dst.getData(new Float32Array(2)); + expect(first).toBeCloseTo(1); + expect(second).toBeCloseTo(0.2); + } finally { + src.dispose(); + dst.dispose(); + } + }); +}); diff --git a/packages/react-native-executorch/__tests__/support/fakeOps.ts b/packages/react-native-executorch/__tests__/support/fakeOps.ts index 4f69fd5ab7..7c653bf240 100644 --- a/packages/react-native-executorch/__tests__/support/fakeOps.ts +++ b/packages/react-native-executorch/__tests__/support/fakeOps.ts @@ -299,14 +299,14 @@ export const cv = { opts: { alpha?: number | readonly number[]; beta?: number | readonly number[] } ): FakeTensor { expectShape(dst, src.shape, 'normalize: dst'); + // Native requires both coefficients, so the TS wrapper must resolve them. + if (opts.alpha === undefined || opts.beta === undefined) { + throw new Error('normalize: alpha and beta are required'); + } const channels = src.shape[0]!; const perChannel = src.numel / channels; - const coefficient = ( - value: number | readonly number[] | undefined, - c: number, - fallback: number - ) => - value === undefined ? fallback : typeof value === 'number' ? value : (value[c] ?? fallback); + const coefficient = (value: number | readonly number[], c: number, fallback: number) => + typeof value === 'number' ? value : (value[c] ?? fallback); for (let i = 0; i < src.numel; i++) { const c = Math.floor(i / perChannel); diff --git a/packages/react-native-executorch/__tests__/tasks/llmChatSession.test.ts b/packages/react-native-executorch/__tests__/tasks/llmChatSession.test.ts index e3f1269baa..867ef556b6 100644 --- a/packages/react-native-executorch/__tests__/tasks/llmChatSession.test.ts +++ b/packages/react-native-executorch/__tests__/tasks/llmChatSession.test.ts @@ -162,6 +162,23 @@ describe('createLLMChatSession — construction', () => { expect(promptsSent().join('')).not.toContain(''); }); + it('keeps the session default for a per-turn option passed as undefined', async () => { + writeTokenizerConfig({ + // eslint-disable-next-line camelcase + chat_template: CHAT_TEMPLATE.replace( + '<|assistant|>', + '<|assistant|>{% if enable_thinking is not defined or enable_thinking %}\n{% else %}DIRECT{% endif %}' + ), + }); + const session = tracked( + await createLLMChatSession(config, { generationConfig: { enableThinking: false } }) + ); + + await session.sendMessage('hi', undefined, { enableThinking: undefined }); + + expect(promptsSent()).toContain('<|assistant|>DIRECT'); + }); + it.each(SPECIAL_TOKEN_CONFIGS)( 'passes configured special tokens to the template (%#)', async (tokens) => { diff --git a/packages/react-native-executorch/__tests__/tasks/remainingTasks.test.ts b/packages/react-native-executorch/__tests__/tasks/remainingTasks.test.ts index 454e3fc245..96f8008a2c 100644 --- a/packages/react-native-executorch/__tests__/tasks/remainingTasks.test.ts +++ b/packages/react-native-executorch/__tests__/tasks/remainingTasks.test.ts @@ -333,6 +333,26 @@ describe('createFsmnVoiceActivityDetector', () => { expect(await vad.detectVoice(new Float32Array(100))).toEqual([]); vad.dispose(); }); + + it('falls back to the default options for fields passed as undefined', async () => { + // Every frame scores 0 non-speech probability, so the defaults find speech. + fakeJsi.registerModel(VAD_PATH, { schema: VAD_SCHEMA }); + + const vad = await createFsmnVoiceActivityDetector(VAD_CONFIG); + const waveform = new Float32Array(16000); + const expected = await vad.detectVoice(waveform); + expect(expected.length).toBeGreaterThan(0); + expect( + await vad.detectVoice(waveform, { + speechThreshold: undefined, + minSpeechDurationMs: undefined, + minSilenceDurationMs: undefined, + speechPadMs: undefined, + mergeGapMs: undefined, + }) + ).toEqual(expected); + vad.dispose(); + }); }); // ============================================================================ diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts index c8baa2873b..85465022b6 100644 --- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts +++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts @@ -228,11 +228,10 @@ export function toChannelsLast(src: Tensor, dst: Tensor): Tensor { */ export function normalize(src: Tensor, dst: Tensor, options?: NormalizeOptions): Tensor { 'worklet'; - const defaultNormalizeOptions = { - alpha: 1 / 255.0, - beta: 0.0, - } as const; - return rnexecutorchJsi.cv.normalize(src, dst, { ...defaultNormalizeOptions, ...options }); + return rnexecutorchJsi.cv.normalize(src, dst, { + alpha: options?.alpha ?? 1 / 255.0, + beta: options?.beta ?? 0.0, + }); } /** diff --git a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts index d9dd5955da..ebd7293a40 100644 --- a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts +++ b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts @@ -256,7 +256,11 @@ export async function createLLMChatSession( throw RnExecuTorchError('RESOURCE_BUSY', 'sendMessage: A turn is already in progress.'); } - const turnGenConfig = { ...defaultGenerationConfig, ...genConfig }; + // Drop fields passed as `undefined` so they keep the session default. + const turnOverrides = Object.fromEntries( + Object.entries(genConfig ?? {}).filter(([, value]) => value !== undefined) + ); + const turnGenConfig: LLMGenerationConfig = { ...defaultGenerationConfig, ...turnOverrides }; const prefillOptions = { enableThinking: turnGenConfig.enableThinking, addGenPrompt: false }; const renderOptions = { ...prefillOptions, addGenPrompt: true }; const generationOpts = { genConfig: turnGenConfig, stopTokens, stopRegex, onToken }; diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts index 667658c56d..4142db4338 100644 --- a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts +++ b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts @@ -276,7 +276,13 @@ export async function createFsmnVoiceActivityDetector( const detectVoiceWorklet = (waveform: Float32Array, options?: VadOptions): VadSegment[] => { 'worklet'; - const mergedOpts: Required = { ...defaultOptions, ...options }; + const mergedOpts: Required = { + speechThreshold: options?.speechThreshold ?? defaultOptions.speechThreshold, + minSpeechDurationMs: options?.minSpeechDurationMs ?? defaultOptions.minSpeechDurationMs, + minSilenceDurationMs: options?.minSilenceDurationMs ?? defaultOptions.minSilenceDurationMs, + speechPadMs: options?.speechPadMs ?? defaultOptions.speechPadMs, + mergeGapMs: options?.mergeGapMs ?? defaultOptions.mergeGapMs, + }; const numFrames = Math.floor((waveform.length - FRAME_LENGTH) / HOP_LENGTH); if (numFrames <= 0) return [];