@tduncan / Cortana / commits / aff592d

feat: implement apple_mlx package and integrate OnDeviceMlxService into chat providers

abdulmominsakib committed Aug 26, 2026 at 15:15 UTC aff592df2dff2120eb3b5f3f1ce179010b0c38cf

28 files changed +1622 -103

lib/features/chat/providers/chat_notifier.dart

@@ -1711,6 +1711,7 @@ class ChatNotifier extends Notifier<ChatState> {

1711 dio: ref.read(dioProvider),

1712 onDeviceGemmaService: ref.read(onDeviceGemmaServiceProvider),

1713 onDeviceLlamaService: ref.read(onDeviceLlamaServiceProvider),

1714 + onDeviceMlxService: ref.read(onDeviceMlxServiceProvider),

1715 loadedOnDeviceRuntime: ref.read(

1716 onDeviceEngineProvider.select((s) => s.loadedRuntime),

1717 ),

lib/features/chat/providers/chat_service_providers.dart

@@ -11,6 +11,7 @@ import 'package:localmind/features/on_device/data/on_device_chat_service.dart';

11 import 'package:localmind/features/on_device/data/on_device_gemma_service.dart';

12 import 'package:localmind/features/on_device/data/on_device_llama_service.dart';

13 import 'package:localmind/features/on_device/data/models/on_device_model.dart';

14 +import 'package:localmind/features/on_device/data/on_device_mlx_service.dart';

15 import '../data/chat_service.dart';

16 import '../data/smart_reply_service.dart';

17 import '../data/title_generation_service.dart';

@@ -33,6 +34,7 @@ final chatServiceProvider = Provider<ChatService?>((ref) {

34 dio: ref.read(dioProvider),

35 onDeviceGemmaService: ref.read(onDeviceGemmaServiceProvider),

36 onDeviceLlamaService: ref.read(onDeviceLlamaServiceProvider),

37 + onDeviceMlxService: ref.read(onDeviceMlxServiceProvider),

38 loadedOnDeviceRuntime: loadedRuntime,

39 imageCompressionEnabled: settings.$1,

40 imageCompressionLevel: settings.$2,

@@ -53,6 +55,7 @@ ChatService createChatServiceForServer({

55 required Dio dio,

56 required OnDeviceGemmaService onDeviceGemmaService,

57 required OnDeviceLlamaService onDeviceLlamaService,

58 + OnDeviceMlxService? onDeviceMlxService,

59 OnDeviceModelRuntime? loadedOnDeviceRuntime,

60 bool imageCompressionEnabled = true,

61 ImageCompressionLevel imageCompressionLevel = ImageCompressionLevel.medium,

@@ -61,6 +64,10 @@ ChatService createChatServiceForServer({

64 if (loadedOnDeviceRuntime == OnDeviceModelRuntime.llamaCpp) {

65 return OnDeviceLlamaChatService(onDeviceLlamaService);

66 }

67 + if (loadedOnDeviceRuntime == OnDeviceModelRuntime.mlx &&

68 + onDeviceMlxService != null) {

69 + return OnDeviceMlxChatService(onDeviceMlxService);

70 + }

72 return ChatService.forServer(

73 server.type,

lib/features/on_device/components/model_card.dart

+22 -1

@@ -166,6 +166,22 @@ class _ModelCardHeader extends StatelessWidget {

166 ),

167 ),

168 ),

169 + if (model.isMlx)

170 + Container(

171 + margin: const EdgeInsets.only(left: 6),

172 + padding: const EdgeInsets.symmetric(horizontal: 8, vertical: 2),

173 + decoration: BoxDecoration(

174 + color: Colors.purple.withValues(alpha: 0.1),

175 + borderRadius: BorderRadius.circular(4),

176 + ),

177 + child: Text(

178 + ' Apple MLX',

179 + style: theme.textTheme.labelSmall?.copyWith(

180 + color: Colors.purple,

181 + fontWeight: FontWeight.bold,

182 + ),

183 + ),

184 + ),

185 ],

186 );

187 }

@@ -218,12 +234,14 @@ class _CapabilityChips extends StatelessWidget {

234 final chips = <String>[

235 if (model.format == OnDeviceModelFormat.gguf) 'GGUF',

236 if (model.runtime == OnDeviceModelRuntime.llamaCpp) 'llama.cpp',

237 + if (model.runtime == OnDeviceModelRuntime.mlx) ' Metal GPU',

238 if (model.supportsFunctionCalling) 'Tools',

239 if (model.supportsThinking) 'Thinking',

240 if (model.supportsVision) 'Vision',

241 model.languagesLabel,

242 if (model.backendNote != null &&

226 - model.backendNote!.toLowerCase() != 'llama.cpp')

243 + model.backendNote!.toLowerCase() != 'llama.cpp' &&

244 + model.backendNote!.toLowerCase() != 'apple mlx')

245 model.backendNote!,

246 if (model.requiresHuggingFaceToken) l10n.model_requires_huggingface_token,

247 ];

@@ -519,6 +537,9 @@ class _DownloadedActions extends ConsumerWidget {

537 await ref

538 .read(importedGgufModelsProvider.notifier)

539 .deleteModel(model.id);

540 + } else if (model.runtime == OnDeviceModelRuntime.mlx) {

541 + final mlxService = ref.read(onDeviceMlxServiceProvider);

542 + await mlxService.deleteModel(model.id);

543 } else {

544 final gemmaService = ref.read(onDeviceGemmaServiceProvider);

545 await gemmaService.deleteModel(model.id);

lib/features/on_device/components/on_device_picker_section.dart

+24 -2

@@ -327,6 +327,26 @@ class _OnDeviceModelTile extends ConsumerWidget {

327 ),

328 ),

329 ),

330 + if (model.isMlx)

331 + Container(

332 + margin: const EdgeInsets.only(left: 6),

333 + padding: const EdgeInsets.symmetric(

334 + horizontal: 6,

335 + vertical: 2,

336 + ),

337 + decoration: BoxDecoration(

338 + color: Colors.purple.withValues(alpha: 0.1),

339 + borderRadius: BorderRadius.circular(4),

340 + ),

341 + child: const Text(

342 + ' MLX',

343 + style: TextStyle(

344 + fontSize: 9,

345 + fontWeight: FontWeight.bold,

346 + color: Colors.purple,

347 + ),

348 + ),

349 + ),

350 ],

351 ),

352 if (!Platform.isIOS &&

@@ -384,9 +404,9 @@ class _OnDeviceModelTile extends ConsumerWidget {

404 ],

405 ),

406 ),

387 - const SizedBox(height: 4),

407 + const SizedBox(height: 6),

408 Wrap(

389 - spacing: 6,

409 + spacing: 4,

410 runSpacing: 4,

411 children: [

412 MetadataChip(

@@ -400,6 +420,8 @@ class _OnDeviceModelTile extends ConsumerWidget {

420 ),

421 if (model.runtime == OnDeviceModelRuntime.llamaCpp)

422 MetadataChip(label: 'llama.cpp', isDark: isDark),

423 + if (model.runtime == OnDeviceModelRuntime.mlx)

424 + MetadataChip(label: ' Apple MLX', isDark: isDark),

425 if (!model.isImported)

426 MetadataChip(label: model.license, isDark: isDark),

427 if (!model.isImported)

lib/features/on_device/data/models/on_device_model.dart

+77 -2

@@ -2,9 +2,9 @@ import 'package:flutter_gemma/flutter_gemma.dart';

3 import '../../../../core/models/enums.dart';

5 -enum OnDeviceModelRuntime { gemma, llamaCpp }

5 +enum OnDeviceModelRuntime { gemma, llamaCpp, mlx }

7 -enum OnDeviceModelFormat { litertlm, task, binary, gguf }

7 +enum OnDeviceModelFormat { litertlm, task, binary, gguf, mlx }

9 enum OnDeviceImportedSource { localFile, huggingFace }

@@ -76,6 +76,7 @@ class OnDeviceModel {

76 }

78 bool get isLlamaCpp => runtime == OnDeviceModelRuntime.llamaCpp;

79 + bool get isMlx => runtime == OnDeviceModelRuntime.mlx;

81 bool get isImportedFromLocalFile =>

82 importedSource == OnDeviceImportedSource.localFile;

@@ -311,6 +312,80 @@ class OnDeviceModel {

312 isCpuOnly: true,

313 ),

314 ];

315 +

316 + static const List<OnDeviceModel> curatedMlxModels = [

317 + OnDeviceModel(

318 + id: 'mlx-qwen-2.5-0.5b',

319 + name: 'Qwen 2.5 0.5B (MLX)',

320 + huggingFaceUrl:

321 + 'https://huggingface.co/mlx-community/Qwen2.5-0.5B-Instruct-4bit',

322 + fileSizeBytes: 408944640,

323 + license: 'Apache-2.0',

324 + description:

325 + 'High efficiency compact Qwen 2.5 4-bit model running via Apple MLX.',

326 + isRecommended: true,

327 + minRamMb: 1536,

328 + parameterLabel: '0.5B (4-bit)',

329 + bestFor: 'Compact multilingual chat with tiny footprint',

330 + languagesLabel: 'Multilingual',

331 + backendNote: 'Apple MLX',

332 + runtime: OnDeviceModelRuntime.mlx,

333 + format: OnDeviceModelFormat.mlx,

334 + ),

335 + OnDeviceModel(

336 + id: 'mlx-llama-3.2-1b',

337 + name: 'Llama 3.2 1B (MLX)',

338 + huggingFaceUrl:

339 + 'https://huggingface.co/mlx-community/Llama-3.2-1B-Instruct-4bit',

340 + fileSizeBytes: 786432000,

341 + license: 'Llama-3.2',

342 + description:

343 + 'Compact, ultra-fast Llama 3.2 4-bit model powered by Apple MLX Metal kernels.',

344 + isRecommended: true,

345 + minRamMb: 2048,

346 + parameterLabel: '1B (4-bit)',

347 + bestFor: 'Lightweight high-speed text generation',

348 + supportsThinking: false,

349 + languagesLabel: 'Multilingual',

350 + backendNote: 'Apple MLX',

351 + runtime: OnDeviceModelRuntime.mlx,

352 + format: OnDeviceModelFormat.mlx,

353 + ),

354 + OnDeviceModel(

355 + id: 'mlx-gemma-2-2b',

356 + name: 'Gemma 2 2B (MLX)',

357 + huggingFaceUrl: 'https://huggingface.co/mlx-community/gemma-2-2b-it-4bit',

358 + fileSizeBytes: 1530920960,

359 + license: 'Gemma',

360 + description:

361 + 'Native Apple MLX 4-bit model running with Metal GPU acceleration.',

362 + minRamMb: 4096,

363 + parameterLabel: '2B (4-bit)',

364 + bestFor: 'Fast chat on Apple Silicon',

365 + supportsThinking: false,

366 + languagesLabel: 'Multilingual',

367 + backendNote: 'Apple MLX',

368 + runtime: OnDeviceModelRuntime.mlx,

369 + format: OnDeviceModelFormat.mlx,

370 + ),

371 + OnDeviceModel(

372 + id: 'mlx-smollm2-135m',

373 + name: 'SmolLM2 135M (MLX)',

374 + huggingFaceUrl:

375 + 'https://huggingface.co/mlx-community/SmolLM2-135M-Instruct',

376 + fileSizeBytes: 272629760,

377 + license: 'Apache-2.0',

378 + description:

379 + 'Ultra-compact 135M MLX model for near-instant on-device inference.',

380 + minRamMb: 1024,

381 + parameterLabel: '135M',

382 + bestFor: 'Instant local inference and memory efficiency',

383 + languagesLabel: 'English',

384 + backendNote: 'Apple MLX',

385 + runtime: OnDeviceModelRuntime.mlx,

386 + format: OnDeviceModelFormat.mlx,

387 + ),

388 + ];

389 }

390

391 class DownloadedModel {

lib/features/on_device/data/on_device_mlx_service.dart new

+415

@@ -0,0 +1,415 @@

1 +import 'dart:async';

2 +import 'dart:io';

3 +

4 +import 'package:apple_mlx/apple_mlx.dart';

5 +import 'package:dio/dio.dart';

6 +import 'package:path/path.dart' as p;

7 +import 'package:path_provider/path_provider.dart';

8 +

9 +import '../../../core/logger/app_logger.dart';

10 +import '../../../core/models/enums.dart';

11 +import '../../servers/data/models/server.dart';

12 +import '../../chat/data/models/chat_parameters.dart';

13 +import '../../chat/data/models/message.dart';

14 +import '../../chat/data/models/mcp_integration.dart';

15 +import '../../chat/data/tools/tool_definition.dart';

16 +import '../../chat/data/chat_service.dart';

17 +import 'models/on_device_model.dart';

18 +

19 +class OnDeviceMlxService {

20 + OnDeviceMlxService(this._dio);

21 +

22 + final Dio _dio;

23 + AppleMlxLlmManager? _llmManager;

24 + String? _currentModelId;

25 + bool _isDisposed = false;

26 + final Map<String, CancelToken> _activeDownloads = {};

27 +

28 + String? get currentModelId => _currentModelId;

29 + bool get isLoaded => _llmManager?.isModelLoaded == true && !_isDisposed;

30 +

31 + /// Returns the base directory where MLX models are stored.

32 + Future<Directory> get _modelsDirectory async {

33 + final appDir = await getApplicationDocumentsDirectory();

34 + final mlxDir = Directory(p.join(appDir.path, 'models', 'mlx'));

35 + if (!await mlxDir.exists()) {

36 + await mlxDir.create(recursive: true);

37 + }

38 + return mlxDir;

39 + }

40 +

41 + /// Returns the specific directory for a model.

42 + Future<Directory> getModelDirectory(String modelId) async {

43 + final base = await _modelsDirectory;

44 + final sanitized = modelId.replaceAll('/', '_');

45 + return Directory(p.join(base.path, sanitized));

46 + }

47 +

48 + /// Checks if a model is fully downloaded and installed.

49 + Future<bool> isModelInstalled(String modelId) async {

50 + try {

51 + final dir = await getModelDirectory(modelId);

52 + if (!await dir.exists()) return false;

53 +

54 + final config = File(p.join(dir.path, 'config.json'));

55 + if (!await config.exists()) return false;

56 +

57 + final files = await dir.list().toList();

58 + final hasWeights = files.any(

59 + (f) =>

60 + f.path.endsWith('.safetensors') ||

61 + f.path.endsWith('.npz') ||

62 + f.path.endsWith('.bin'),

63 + );

64 + return hasWeights;

65 + } catch (e) {

66 + Log.error('Error checking MLX model installation for $modelId: $e');

67 + return false;

68 + }

69 + }

70 +

71 + /// Returns a set of installed MLX model IDs.

72 + Future<Set<String>> getInstalledModelIds() async {

73 + if (!Platform.isIOS && !Platform.isMacOS) {

74 + return const <String>{};

75 + }

76 + try {

77 + final base = await _modelsDirectory;

78 + if (!await base.exists()) return const <String>{};

79 +

80 + final installed = <String>{};

81 + final curated = OnDeviceModel.curatedMlxModels;

82 +

83 + for (final model in curated) {

84 + if (await isModelInstalled(model.id)) {

85 + installed.add(model.id);

86 + }

87 + }

88 + return installed;

89 + } catch (e) {

90 + Log.error('Error listing installed MLX models: $e');

91 + return const <String>{};

92 + }

93 + }

94 +

95 + /// Downloads all required files of an MLX model from Hugging Face.

96 + Future<void> downloadModel(

97 + OnDeviceModel model, {

98 + String? huggingFaceToken,

99 + void Function(int receivedBytes, int totalBytes)? onProgress,

100 + CancelToken? cancelToken,

101 + }) async {

102 + if (!Platform.isIOS && !Platform.isMacOS) {

103 + throw UnsupportedError(

104 + 'Apple MLX models are only supported on Apple Silicon (iOS/macOS).',

105 + );

106 + }

107 +

108 + final targetDir = await getModelDirectory(model.id);

109 + if (!await targetDir.exists()) {

110 + await targetDir.create(recursive: true);

111 + }

112 +

113 + // Extract repo ID from huggingFaceUrl if needed

114 + final repoId = _extractRepoId(model.huggingFaceUrl);

115 + final effectiveCancelToken = cancelToken ?? CancelToken();

116 + _activeDownloads[model.id] = effectiveCancelToken;

117 +

118 + try {

119 + Log.info('Fetching MLX model file list for $repoId from Hugging Face...');

120 +

121 + final headers = <String, dynamic>{

122 + if (huggingFaceToken != null && huggingFaceToken.isNotEmpty)

123 + 'Authorization': 'Bearer $huggingFaceToken',

124 + };

125 +

126 + // 1. Fetch file list from Hugging Face API

127 + final apiUrl = 'https://huggingface.co/api/models/$repoId/tree/main';

128 + List<String> filesToDownload = [];

129 + int totalModelBytes = model.fileSizeBytes;

130 +

131 + try {

132 + final res = await _dio.get<List<dynamic>>(

133 + apiUrl,

134 + options: Options(headers: headers),

135 + cancelToken: effectiveCancelToken,

136 + );

137 + if (res.data != null) {

138 + int computedTotal = 0;

139 + for (final item in res.data!) {

140 + if (item is Map) {

141 + final path = item['path'] as String?;

142 + final type = item['type'] as String?;

143 + final size = (item['size'] as num?)?.toInt() ?? 0;

144 + if (type == 'file' && path != null && !_shouldSkipFile(path)) {

145 + filesToDownload.add(path);

146 + computedTotal += size;

147 + }

148 + }

149 + }

150 + if (computedTotal > 0) {

151 + totalModelBytes = computedTotal;

152 + }

153 + }

154 + } catch (e) {

155 + Log.warning(

156 + 'Could not fetch Hugging Face file tree ($e). Using default MLX file list.',

157 + );

158 + filesToDownload = [

159 + 'config.json',

160 + 'model.safetensors',

161 + 'tokenizer.json',

162 + 'tokenizer_config.json',

163 + 'special_tokens_map.json',

164 + ];

165 + }

166 +

167 + if (filesToDownload.isEmpty) {

168 + filesToDownload = [

169 + 'config.json',

170 + 'model.safetensors',

171 + 'tokenizer.json',

172 + 'tokenizer_config.json',

173 + ];

174 + }

175 +

176 + int cumulativeDownloaded = 0;

177 + final fileSizes = <String, int>{};

178 +

179 + for (final fileName in filesToDownload) {

180 + if (effectiveCancelToken.isCancelled) break;

181 +

182 + final fileUrl = 'https://huggingface.co/$repoId/resolve/main/$fileName';

183 + final savePath = p.join(targetDir.path, fileName);

184 + int fileDownloadedBefore = fileSizes[fileName] ?? 0;

185 +

186 + await _dio.download(

187 + fileUrl,

188 + savePath,

189 + options: Options(headers: headers),

190 + cancelToken: effectiveCancelToken,

191 + onReceiveProgress: (received, total) {

192 + final delta = received - fileDownloadedBefore;

193 + fileDownloadedBefore = received;

194 + fileSizes[fileName] = received;

195 + cumulativeDownloaded += delta;

196 +

197 + onProgress?.call(

198 + cumulativeDownloaded,

199 + totalModelBytes > 0 ? totalModelBytes : cumulativeDownloaded,

200 + );

201 + },

202 + );

203 + }

204 +

205 + Log.info(

206 + 'MLX model ${model.id} successfully downloaded to ${targetDir.path}',

207 + );

208 + } catch (e) {

209 + Log.error('Failed to download MLX model ${model.id}: $e');

210 + rethrow;

211 + } finally {

212 + _activeDownloads.remove(model.id);

213 + }

214 + }

215 +

216 + /// Cancels an in-flight download.

217 + void cancelDownload(String modelId) {

218 + _activeDownloads[modelId]?.cancel('Download cancelled by user');

219 + _activeDownloads.remove(modelId);

220 + }

221 +

222 + /// Deletes a downloaded model from disk.

223 + Future<void> deleteModel(String modelId) async {

224 + try {

225 + if (_currentModelId == modelId) {

226 + await unloadModel();

227 + }

228 + final dir = await getModelDirectory(modelId);

229 + if (await dir.exists()) {

230 + await dir.delete(recursive: true);

231 + }

232 + Log.info('Deleted MLX model $modelId');

233 + } catch (e) {

234 + Log.error('Error deleting MLX model $modelId: $e');

235 + }

236 + }

237 +

238 + /// Loads an MLX model for inference.

239 + Future<void> loadModel(OnDeviceModel model) async {

240 + if (_isDisposed) {

241 + throw StateError('OnDeviceMlxService has been disposed');

242 + }

243 +

244 + final dir = await getModelDirectory(model.id);

245 + if (!await isModelInstalled(model.id)) {

246 + throw StateError(

247 + 'MLX Model ${model.id} is not downloaded. Please download it first.',

248 + );

249 + }

250 +

251 + await unloadModel();

252 +

253 + Log.info('Loading MLX model ${model.id} from ${dir.path}');

254 + final manager = AppleMlxLlmManager();

255 + await manager.loadAndStartServer(

256 + modelPath: dir.path,

257 + modelId: _extractRepoId(model.huggingFaceUrl),

258 + thinkingEnabled: model.supportsThinking,

259 + );

260 +

261 + _llmManager = manager;

262 + _currentModelId = model.id;

263 + Log.info('MLX model ${model.id} loaded successfully');

264 + }

265 +

266 + /// Unloads the active MLX model.

267 + Future<void> unloadModel() async {

268 + if (_llmManager != null) {

269 + Log.info('Unloading MLX model $_currentModelId');

270 + await _llmManager!.dispose();

271 + _llmManager = null;

272 + _currentModelId = null;

273 + }

274 + }

275 +

276 + /// Streams chat responses from the loaded MLX model.

277 + Stream<ChatResponse> sendMessage({

278 + required String modelId,

279 + required List<Message> messages,

280 + required ChatParameters params,

281 + }) async* {

282 + final manager = _llmManager;

283 + if (manager == null || !manager.isModelLoaded) {

284 + yield const ChatResponse(

285 + type: ChatResponseType.error,

286 + content: 'MLX model is not loaded. Please load the model first.',

287 + );

288 + yield const ChatResponse(type: ChatResponseType.done);

289 + return;

290 + }

291 +

292 + final formattedMessages = <Map<String, String>>[];

293 +

294 + if (params.systemPrompt != null && params.systemPrompt!.isNotEmpty) {

295 + formattedMessages.add({

296 + 'role': 'system',

297 + 'content': params.systemPrompt!,

298 + });

299 + }

300 +

301 + for (final m in messages) {

302 + final role = m.role == MessageRole.user

303 + ? 'user'

304 + : m.role == MessageRole.assistant

305 + ? 'assistant'

306 + : 'system';

307 + formattedMessages.add({'role': role, 'content': m.content});

308 + }

309 +

310 + try {

311 + final stream = manager.generateStream(

312 + messages: formattedMessages,

313 + temperature: params.temperature,

314 + maxTokens: params.maxTokens,

315 + thinking: true,

316 + );

317 +

318 + await for (final delta in stream) {

319 + if (delta.isDone) {

320 + yield const ChatResponse(type: ChatResponseType.done);

321 + break;

322 + }

323 +

324 + if (delta.reasoningContent != null &&

325 + delta.reasoningContent!.isNotEmpty) {

326 + yield ChatResponse(

327 + type: ChatResponseType.reasoning,

328 + content: delta.reasoningContent!,

329 + );

330 + }

331 +

332 + if (delta.content != null && delta.content!.isNotEmpty) {

333 + yield ChatResponse(

334 + type: ChatResponseType.message,

335 + content: delta.content!,

336 + );

337 + }

338 + }

339 + } catch (e) {

340 + Log.error('MLX generation error: $e');

341 + yield ChatResponse(

342 + type: ChatResponseType.error,

343 + content: 'MLX generation error: $e',

344 + );

345 + yield const ChatResponse(type: ChatResponseType.done);

346 + }

347 + }

348 +

349 + void dispose() {

350 + _isDisposed = true;

351 + for (final token in _activeDownloads.values) {

352 + token.cancel('Service disposed');

353 + }

354 + _activeDownloads.clear();

355 + unawaited(unloadModel());

356 + }

357 +

358 + String _extractRepoId(String url) {

359 + if (!url.startsWith('http')) return url;

360 + final uri = Uri.tryParse(url);

361 + if (uri == null) return url;

362 + final segments = uri.pathSegments;

363 + if (segments.length >= 2) {

364 + return '${segments[0]}/${segments[1]}';

365 + }

366 + return url;

367 + }

368 +

369 + bool _shouldSkipFile(String path) {

370 + final lower = path.toLowerCase();

371 + return lower.startsWith('.') ||

372 + lower.endsWith('.md') ||

373 + lower.endsWith('.gitattributes') ||

374 + lower.endsWith('.png') ||

375 + lower.endsWith('.jpg') ||

376 + lower.endsWith('.jpeg');

377 + }

378 +}

379 +

380 +class OnDeviceMlxChatService implements ChatService {

381 + OnDeviceMlxChatService(this._mlxService);

382 +

383 + final OnDeviceMlxService _mlxService;

384 + bool _isCancelled = false;

385 +

386 + @override

387 + Stream<ChatResponse> sendMessage({

388 + required Server server,

389 + required String modelId,

390 + required List<Message> messages,

391 + required ChatParameters params,

392 + List<McpIntegration>? integrations,

393 + List<ToolDefinition>? tools,

394 + bool continueGeneration = false,

395 + }) async* {

396 + _isCancelled = false;

397 + await for (final response in _mlxService.sendMessage(

398 + modelId: modelId,

399 + messages: messages,

400 + params: params,

401 + )) {

402 + if (_isCancelled) break;

403 + yield response;

404 + if (response.type == ChatResponseType.done ||

405 + response.type == ChatResponseType.error) {

406 + break;

407 + }

408 + }

409 + }

410 +

411 + @override

412 + void cancelStream() {

413 + _isCancelled = true;

414 + }

415 +}

lib/features/on_device/providers/foreground_download_providers.dart

+74 -31

@@ -9,6 +9,7 @@ import '../../../core/providers/review_prompt_providers.dart';

9 import '../../../core/providers/storage_providers.dart';

10 import '../data/models/download_progress_info.dart';

11 import '../data/models/download_status.dart';

12 +import '../data/models/on_device_model.dart';

13 import 'on_device_providers.dart';

15 final foregroundDownloadNotifierProvider =

@@ -92,42 +93,82 @@ class ForegroundDownloadNotifier

93 _saveState();

95 final gemmaService = ref.read(onDeviceGemmaServiceProvider);

96 + final mlxService = ref.read(onDeviceMlxServiceProvider);

97 final hfToken = ref.read(

98 settingsProvider.select((s) => s.huggingFaceToken),

99 );

100

101 try {

100 - await gemmaService.installModel(

101 - model,

102 - token: (hfToken != null && hfToken.isNotEmpty) ? hfToken : null,

103 - onProgress: (progress) {

104 - if (!ref.mounted || _activeDownloads[modelId] != true) return;

105 -

106 - final startTime = _downloadStartTimes[modelId] ?? DateTime.now();

107 - final elapsed = DateTime.now().difference(startTime).inSeconds;

108 - final total = model.fileSizeBytes;

109 - final received = ((progress / 100.0) * total).round();

110 -

111 - final bytesPerSecond = elapsed > 0 ? (received / elapsed).round() : 0;

112 - final etaSeconds = bytesPerSecond > 0

113 - ? ((total - received) / bytesPerSecond).round()

114 - : null;

115 -

116 - state = {

117 - ...state,

118 - modelId: DownloadProgressInfo(

119 - modelId: modelId,

120 - status: DownloadStatus.running,

121 - progress: progress / 100.0,

122 - receivedBytes: received,

123 - totalBytes: total,

124 - bytesPerSecond: bytesPerSecond,

125 - etaSeconds: etaSeconds,

126 - ),

127 - };

128 - _saveState();

129 - },

130 - );

102 + if (model.runtime == OnDeviceModelRuntime.mlx) {

103 + await mlxService.downloadModel(

104 + model,

105 + huggingFaceToken: (hfToken != null && hfToken.isNotEmpty)

106 + ? hfToken

107 + : null,

108 + onProgress: (received, total) {

109 + if (!ref.mounted || _activeDownloads[modelId] != true) return;

110 +

111 + final startTime = _downloadStartTimes[modelId] ?? DateTime.now();

112 + final elapsed = DateTime.now().difference(startTime).inSeconds;

113 + final bytesPerSecond = elapsed > 0

114 + ? (received / elapsed).round()

115 + : 0;

116 + final etaSeconds = bytesPerSecond > 0

117 + ? ((total - received) / bytesPerSecond).round()

118 + : null;

119 +

120 + final progress = total > 0 ? (received / total) : 0.0;

121 +

122 + state = {

123 + ...state,

124 + modelId: DownloadProgressInfo(

125 + modelId: modelId,

126 + status: DownloadStatus.running,

127 + progress: progress.clamp(0.0, 1.0),

128 + receivedBytes: received,

129 + totalBytes: total,

130 + bytesPerSecond: bytesPerSecond,

131 + etaSeconds: etaSeconds,

132 + ),

133 + };

134 + _saveState();

135 + },

136 + );

137 + } else {

138 + await gemmaService.installModel(

139 + model,

140 + token: (hfToken != null && hfToken.isNotEmpty) ? hfToken : null,

141 + onProgress: (progress) {

142 + if (!ref.mounted || _activeDownloads[modelId] != true) return;

143 +

144 + final startTime = _downloadStartTimes[modelId] ?? DateTime.now();

145 + final elapsed = DateTime.now().difference(startTime).inSeconds;

146 + final total = model.fileSizeBytes;

147 + final received = ((progress / 100.0) * total).round();

148 +

149 + final bytesPerSecond = elapsed > 0

150 + ? (received / elapsed).round()

151 + : 0;

152 + final etaSeconds = bytesPerSecond > 0

153 + ? ((total - received) / bytesPerSecond).round()

154 + : null;

155 +

156 + state = {

157 + ...state,

158 + modelId: DownloadProgressInfo(

159 + modelId: modelId,

160 + status: DownloadStatus.running,

161 + progress: progress / 100.0,

162 + receivedBytes: received,

163 + totalBytes: total,

164 + bytesPerSecond: bytesPerSecond,

165 + etaSeconds: etaSeconds,

166 + ),

167 + };

168 + _saveState();

169 + },

170 + );

171 + }

172

173 // Download completed

174 if (_activeDownloads[modelId] == true) {

@@ -187,6 +228,8 @@ class ForegroundDownloadNotifier

228 }

229

230 Future<void> cancelDownload(String modelId) async {

231 + final mlxService = ref.read(onDeviceMlxServiceProvider);

232 + mlxService.cancelDownload(modelId);

233 _activeDownloads.remove(modelId);

234 _downloadStartTimes.remove(modelId);

235 state = {...state}..remove(modelId);

lib/features/on_device/providers/on_device_providers.dart

+54 -6

@@ -16,6 +16,7 @@ import '../data/models/on_device_model.dart';

16 import '../data/notification_permission_service.dart';

17 import '../data/on_device_gemma_service.dart';

18 import '../data/on_device_llama_service.dart';

19 +import '../data/on_device_mlx_service.dart';

20 import '../../servers/providers/server_providers.dart';

22 final notificationPermissionServiceProvider =

@@ -35,6 +36,12 @@ final onDeviceLlamaServiceProvider = Provider<OnDeviceLlamaService>((ref) {

36 return service;

37 });

39 +final onDeviceMlxServiceProvider = Provider<OnDeviceMlxService>((ref) {

40 + final service = OnDeviceMlxService(ref.read(dioProvider));

41 + ref.onDispose(() => service.dispose());

42 + return service;

43 +});

44 +

45 final importedGgufModelRepositoryProvider =

46 Provider<ImportedGgufModelRepository>((ref) {

47 return ImportedGgufModelRepository(

@@ -55,14 +62,21 @@ final onDeviceEngineProvider =

63 final onDeviceModelsProvider = Provider<List<OnDeviceModel>>((ref) {

64 final imported = ref.watch(importedGgufModelsProvider);

58 - return [...OnDeviceModel.curatedModels, ...imported];

65 + final mlxModels = Platform.isIOS

66 + ? OnDeviceModel.curatedMlxModels

67 + : const <OnDeviceModel>[];

68 + return [...mlxModels, ...OnDeviceModel.curatedModels, ...imported];

69 });

71 final downloadedModelsProvider = FutureProvider<Set<String>>((ref) async {

72 final gemmaService = ref.read(onDeviceGemmaServiceProvider);

73 final installed = await gemmaService.getInstalledModelIds();

74 + final mlxService = ref.read(onDeviceMlxServiceProvider);

75 + final mlxInstalled = Platform.isIOS

76 + ? await mlxService.getInstalledModelIds()

77 + : const <String>{};

78 final imported = ref.watch(importedGgufModelsProvider);

65 - return {...installed, ...imported.map((model) => model.id)};

79 + return {...installed, ...mlxInstalled, ...imported.map((model) => model.id)};

80 });

82 final onDeviceModelStateProvider =

@@ -239,6 +253,7 @@ class OnDeviceEngineNotifier extends Notifier<OnDeviceEngineState> {

253 ref.read(onDeviceGemmaServiceProvider);

254 OnDeviceLlamaService get _llamaService =>

255 ref.read(onDeviceLlamaServiceProvider);

256 + OnDeviceMlxService get _mlxService => ref.read(onDeviceMlxServiceProvider);

257

258 Future<void> loadModel(String modelId, PreferredBackend backend) async {

259 if (!ref.mounted) return;

@@ -287,6 +302,7 @@ class OnDeviceEngineNotifier extends Notifier<OnDeviceEngineState> {

302

303 final settings = ref.read(settingsProvider);

304 await _gemmaService.unloadModel();

305 + await _mlxService.unloadModel();

306 if (!ref.mounted) return;

307 await _llamaService.loadModel(

308 model,

@@ -294,6 +310,24 @@ class OnDeviceEngineNotifier extends Notifier<OnDeviceEngineState> {

310 useGpu: !model.isCpuOnly && effectiveBackend == PreferredBackend.gpu,

311 );

312 if (!ref.mounted) return;

313 + } else if (model.runtime == OnDeviceModelRuntime.mlx) {

314 + final isInstalled = await _mlxService.isModelInstalled(modelId);

315 + if (!ref.mounted) return;

316 + if (!isInstalled) {

317 + state = state.copyWith(

318 + status: OnDeviceEngineStatus.error,

319 + loadedModelId: null,

320 + loadedRuntime: null,

321 + error: 'MLX Model not found. Please download it first.',

322 + );

323 + return;

324 + }

325 +

326 + await _gemmaService.unloadModel();

327 + await _llamaService.unloadModel();

328 + if (!ref.mounted) return;

329 + await _mlxService.loadModel(model);

330 + if (!ref.mounted) return;

331 } else {

332 final isInstalled = await _gemmaService.isModelInstalled(modelId);

333 if (!ref.mounted) return;

@@ -308,6 +342,7 @@ class OnDeviceEngineNotifier extends Notifier<OnDeviceEngineState> {

342 }

343

344 await _llamaService.unloadModel();

345 + await _mlxService.unloadModel();

346 if (!ref.mounted) return;

347 await _gemmaService.loadModel(modelId, effectiveBackend);

348 if (!ref.mounted) return;

@@ -327,10 +362,14 @@ class OnDeviceEngineNotifier extends Notifier<OnDeviceEngineState> {

362 final onDeviceServer = servers

363 .where((s) => s.type == ServerType.onDevice)

364 .firstOrNull;

330 - if (onDeviceServer != null) {

331 - ref

332 - .read(activeServerIdProvider.notifier)

333 - .setActiveServer(onDeviceServer);

365 + if (onDeviceServer != null && ref.mounted) {

366 + Future.microtask(() {

367 + if (ref.mounted) {

368 + ref

369 + .read(activeServerIdProvider.notifier)

370 + .setActiveServer(onDeviceServer);

371 + }

372 + });

373 }

374 }

375

@@ -364,6 +403,7 @@ class OnDeviceEngineNotifier extends Notifier<OnDeviceEngineState> {

403 Future<void> unloadModel() async {

404 final gemma = ref.mounted ? _gemmaService : null;

405 final llama = ref.mounted ? _llamaService : null;

406 + final mlx = ref.mounted ? _mlxService : null;

407 if (gemma != null) {

408 try {

409 await gemma.unloadModel();

@@ -378,6 +418,13 @@ class OnDeviceEngineNotifier extends Notifier<OnDeviceEngineState> {

418 Log.error('Failed to unload llama.cpp model: $e');

419 }

420 }

421 + if (mlx != null) {

422 + try {

423 + await mlx.unloadModel();

424 + } catch (e) {

425 + Log.error('Failed to unload MLX model: $e');

426 + }

427 + }

428 if (ref.mounted) {

429 state = const OnDeviceEngineState();

430 }

@@ -386,6 +433,7 @@ class OnDeviceEngineNotifier extends Notifier<OnDeviceEngineState> {

433 Future<void> disposeService() async {

434 _gemmaService.dispose();

435 _llamaService.dispose();

436 + _mlxService.dispose();

437 state = const OnDeviceEngineState();

438 }

439 }

lib/features/servers/providers/server_providers.dart

+7 -1

@@ -158,9 +158,15 @@ final availableModelsProvider = FutureProvider.family<List<dynamic>, String>((

158 m.parameterLabel.replaceAll(RegExp(r'[^0-9\.]'), ''),

159 ),

160 fileSize: m.fileSizeBytes,

161 - quantization: m.format == OnDeviceModelFormat.gguf ? 'GGUF' : null,

161 + quantization: m.format == OnDeviceModelFormat.gguf

162 + ? 'GGUF'

163 + : m.format == OnDeviceModelFormat.mlx

164 + ? 'MLX 4-bit'

165 + : null,

166 architecture: m.runtime == OnDeviceModelRuntime.llamaCpp

167 ? 'llama.cpp'

168 + : m.runtime == OnDeviceModelRuntime.mlx

169 + ? 'Apple MLX'

170 : null,

171 serverType: ServerType.onDevice,

172 serverId: server.id,

packages/apple_mlx/README.md new

+88

@@ -0,0 +1,88 @@

1 +# Apple MLX Support (`apple_mlx`)

2 +

3 +Local package integrating Apple's **MLX** ecosystem (`lib_mlx`, `speech_mlx`, `mlx_audio`, and `mlx-c`) with Dart & Flutter to run LLM Text Generation, Speech-to-Text, and Text-to-Speech natively on Apple Silicon (iOS and macOS).

4 +

5 +## Capabilities

6 +

7 +### 1. LLM Text Generation (`AppleMlxLlmManager` & `lib_mlx`)

8 +- **Native Apple Silicon Inference**: Runs quantized MLX models (Gemma, Llama 3.2, Qwen 2.5, SmolLM2) directly on device via Unified Memory and Metal.

9 +- **OpenAI-Compatible Local Endpoint**: Zero-overhead localhost HTTP/SSE streaming client.

10 +- **Reasoning / Thinking Support**: Separates `<thinking>` tokens from final assistant content.

11 +- **Tool Calling Support**: Function and tool calling protocol.

12 +

13 +### 2. Speech-To-Text (`AppleMlxSpeechManager` & `speech_mlx` / `mlx_audio`)

14 +- **Parakeet-TDT & Whisper**: High-accuracy local speech recognition.

15 +- **Isolate Offloading**: Dedicated worker isolates keep audio processing and resampling off the UI isolate.

16 +

17 +### 3. Text-To-Speech (`PocketTTS`)

18 +- **Neural Voice Synthesis**: Real-time voice generation with incremental PCM streaming.

19 +- **Voices**: Built-in multi-voice support (`alba`, `marius`, `javert`, `jean`).

20 +

21 +### 4. Conversational Turn Detection (`SmartTurn`)

22 +- **Smart Turn v3.2**: Sub-millisecond endpoint classifier to determine whether a speaker finished speaking during a voice conversation.

23 +

24 +---

25 +

26 +## Quickstart

27 +

28 +### LLM Text Generation

29 +

30 +```dart

31 +import 'package:apple_mlx/apple_mlx.dart';

32 +

33 +Future<void> main() async {

34 + final llm = AppleMlxLlmManager();

35 +

36 + // 1. Load MLX model and launch local server

37 + final serverInfo = await llm.loadAndStartServer(

38 + modelPath: '/path/to/mlx-community/gemma-4-e2b-it-4bit',

39 + modelId: AppleMlxModelIds.gemma4E2b,

40 + thinkingEnabled: true,

41 + );

42 + print('MLX Server running on ${serverInfo.baseUrl}');

43 +

44 + // 2. Stream generation

45 + final stream = llm.generateStream(

46 + messages: [

47 + {'role': 'user', 'content': 'Explain quantum computing in simple terms.'},

48 + ],

49 + temperature: 0.7,

50 + );

51 +

52 + await for (final delta in stream) {

53 + if (delta.reasoningContent != null) {

54 + print('[Thinking] ${delta.reasoningContent}');

55 + }

56 + if (delta.content != null) {

57 + print(delta.content);

58 + }

59 + }

60 +

61 + // 3. Clean up

62 + await llm.dispose();

63 +}

64 +```

65 +

66 +### STT and TTS Audio

67 +

68 +```dart

69 +import 'package:apple_mlx/apple_mlx.dart';

70 +

71 +Future<void> main() async {

72 + final speech = AppleMlxSpeechManager();

73 + await speech.initialize();

74 +

75 + // Transcribe audio

76 + final text = await speech.transcribeWavFile('input.wav');

77 + print('Transcript: $text');

78 +

79 + // Synthesize voice

80 + await speech.synthesizeToWav(

81 + text: 'Hello from Apple MLX on LocalMind!',

82 + outputPath: 'output.wav',

83 + voice: 'alba',

84 + );

85 +

86 + await speech.dispose();

87 +}

88 +```

packages/apple_mlx/build/native_assets/macos/native_assets.json new

@@ -0,0 +1 @@

1 +{"format-version":[1,0,0],"native-assets":{}}

\ No newline at end of file

packages/apple_mlx/build/test_cache/build/11f9252fab9bb4424d40abfb06fd5e81.cache.dill.track.dill

Binary files /dev/null and b/packages/apple_mlx/build/test_cache/build/11f9252fab9bb4424d40abfb06fd5e81.cache.dill.track.dill differ

packages/apple_mlx/build/unit_test_assets/AssetManifest.bin

Binary files /dev/null and b/packages/apple_mlx/build/unit_test_assets/AssetManifest.bin differ

packages/apple_mlx/build/unit_test_assets/FontManifest.json new

@@ -0,0 +1 @@

1 +[]

\ No newline at end of file

packages/apple_mlx/build/unit_test_assets/NOTICES.Z

Binary files /dev/null and b/packages/apple_mlx/build/unit_test_assets/NOTICES.Z differ

packages/apple_mlx/build/unit_test_assets/NativeAssetsManifest.json new

@@ -0,0 +1 @@

1 +{"format-version":[1,0,0],"native-assets":{}}

\ No newline at end of file

packages/apple_mlx/build/unit_test_assets/shaders/ink_sparkle.frag

Binary files /dev/null and b/packages/apple_mlx/build/unit_test_assets/shaders/ink_sparkle.frag differ

packages/apple_mlx/build/unit_test_assets/shaders/stretch_effect.frag

Binary files /dev/null and b/packages/apple_mlx/build/unit_test_assets/shaders/stretch_effect.frag differ

packages/apple_mlx/lib/apple_mlx.dart new

+13

@@ -0,0 +1,13 @@

1 +library;

2 +

3 +// Core MLX packages

4 +export 'package:lib_mlx/lib_mlx.dart';

5 +export 'package:speech_mlx/speech_mlx.dart';

6 +export 'package:mlx_audio/mlx_audio.dart';

7 +export 'package:speech_core/speech_core.dart';

8 +export 'package:audio_core/audio_core.dart';

9 +

10 +// LocalMind MLX Suite

11 +export 'src/mlx_models.dart';

12 +export 'src/mlx_speech_manager.dart';

13 +export 'src/mlx_llm_manager.dart';

packages/apple_mlx/lib/src/mlx_llm_manager.dart new

+179

@@ -0,0 +1,179 @@

1 +import 'dart:async';

2 +import 'package:lib_mlx/lib_mlx.dart';

3 +

4 +/// Represents streaming chunk delta from an MLX LLM generation.

5 +class AppleMlxChatDelta {

6 + /// Regular content delta chunk.

7 + final String? content;

8 +

9 + /// Reasoning / thinking delta chunk (if model supports thinking).

10 + final String? reasoningContent;

11 +

12 + /// Whether generation is finished.

13 + final bool isDone;

14 +

15 + /// Raw chunk data if available.

16 + final Map<String, Object?>? rawData;

17 +

18 + const AppleMlxChatDelta({

19 + this.content,

20 + this.reasoningContent,

21 + this.isDone = false,

22 + this.rawData,

23 + });

24 +}

25 +

26 +/// High-level manager for running MLX LLM text generation models natively on Apple devices.

27 +class AppleMlxLlmManager {

28 + final LibMlxRuntime _runtime;

29 + LibMlxOpenAiClient? _client;

30 + MlxModelHandle? _handle;

31 + MlxServerInfo? _serverInfo;

32 +

33 + AppleMlxLlmManager({LibMlxRuntime? runtime})

34 + : _runtime = runtime ?? const LibMlxRuntime();

35 +

36 + bool get isModelLoaded => _handle != null;

37 + bool get isServerRunning => _serverInfo != null;

38 + MlxServerInfo? get serverInfo => _serverInfo;

39 + MlxModelHandle? get modelHandle => _handle;

40 +

41 + /// Loads an MLX model from disk and starts the local OpenAI-compatible inference server.

42 + Future<MlxServerInfo> loadAndStartServer({

43 + required String modelPath,

44 + String modelId = MlxAvailableModel.defaultHuggingFaceModelId,

45 + bool thinkingEnabled = true,

46 + bool lazyEncoders = true,

47 + int port = 0,

48 + int queueLimit = 1,

49 + }) async {

50 + // Unload existing session if present

51 + await dispose();

52 +

53 + final config = MlxModelConfig(

54 + modelPath: modelPath,

55 + modelId: modelId,

56 + thinkingEnabled: thinkingEnabled,

57 + lazyEncoders: lazyEncoders,

58 + );

59 +

60 + _handle = await _runtime.loadModel(config);

61 +

62 + final serverConfig = MlxServerConfig(

63 + port: port,

64 + modelId: modelId,

65 + queueLimit: queueLimit,

66 + );

67 +

68 + _serverInfo = await _runtime.startServer(_handle!, config: serverConfig);

69 + _client = LibMlxOpenAiClient(baseUri: _serverInfo!.uri);

70 +

71 + return _serverInfo!;

72 + }

73 +

74 + /// Streams chat completions using Server-Sent Events (SSE).

75 + Stream<AppleMlxChatDelta> generateStream({

76 + required List<Map<String, String>> messages,

77 + double temperature = 0.7,

78 + int? maxTokens,

79 + bool thinking = true,

80 + }) async* {

81 + final client = _client;

82 + final serverInfo = _serverInfo;

83 + if (client == null || serverInfo == null) {

84 + throw StateError(

85 + 'AppleMlxLlmManager: Model is not loaded or server is not running. Call loadAndStartServer() first.',

86 + );

87 + }

88 +

89 + final request = <String, Object?>{

90 + 'model': serverInfo.modelId,

91 + 'messages': messages,

92 + 'temperature': temperature,

93 + 'thinking': thinking,

94 + 'max_tokens': ?maxTokens,

95 + };

96 +

97 + final sseStream = client.chatCompletionsStream(request);

98 +

99 + await for (final event in sseStream) {

100 + if (event.done) {

101 + yield const AppleMlxChatDelta(isDone: true);

102 + break;

103 + }

104 +

105 + final choices = event.data?['choices'] as List?;

106 + final choice = choices?.firstOrNull as Map?;

107 + final delta = choice?['delta'] as Map?;

108 +

109 + if (delta != null) {

110 + final content = delta['content'] as String?;

111 + final reasoning = delta['reasoning_content'] as String?;

112 +

113 + if (content != null || reasoning != null) {

114 + yield AppleMlxChatDelta(

115 + content: content,

116 + reasoningContent: reasoning,

117 + rawData: event.data,

118 + );

119 + }

120 + }

121 + }

122 + }

123 +

124 + /// Generates non-streaming completion response.

125 + Future<String> generate({

126 + required List<Map<String, String>> messages,

127 + double temperature = 0.7,

128 + int? maxTokens,

129 + }) async {

130 + final client = _client;

131 + final serverInfo = _serverInfo;

132 + if (client == null || serverInfo == null) {

133 + throw StateError(

134 + 'AppleMlxLlmManager: Model is not loaded or server is not running. Call loadAndStartServer() first.',

135 + );

136 + }

137 +

138 + final request = <String, Object?>{

139 + 'model': serverInfo.modelId,

140 + 'messages': messages,

141 + 'temperature': temperature,

142 + 'max_tokens': ?maxTokens,

143 + };

144 +

145 + final response = await client.chatCompletions(request);

146 + final choices = response['choices'] as List?;

147 + if (choices != null && choices.isNotEmpty) {

148 + final first = choices.first as Map?;

149 + final message = first?['message'] as Map?;

150 + return message?['content'] as String? ?? '';

151 + }

152 + return '';

153 + }

154 +

155 + /// Checks the current runtime and model status.

156 + Future<MlxRuntimeStatus?> checkStatus() async {

157 + final handle = _handle;

158 + if (handle == null) return null;

159 + return _runtime.serverStatus(handle);

160 + }

161 +

162 + /// Stops server, unloads model, and releases memory.

163 + Future<void> dispose() async {

164 + _client?.close();

165 + _client = null;

166 +

167 + final handle = _handle;

168 + if (handle != null) {

169 + try {

170 + await _runtime.stopServer(handle);

171 + } catch (_) {}

172 + try {

173 + await _runtime.unloadModel(handle);

174 + } catch (_) {}

175 + _handle = null;

176 + }

177 + _serverInfo = null;

178 + }

179 +}

packages/apple_mlx/lib/src/mlx_models.dart new

+52

@@ -0,0 +1,52 @@

1 +/// Known and recommended Apple MLX model checkpoints.

2 +class AppleMlxModelIds {

3 + const AppleMlxModelIds._();

4 +

5 + // LLM / Text Generation Models (Hugging Face mlx-community)

6 + static const String gemma22b = 'mlx-community/gemma-2-2b-it-4bit';

7 + static const String llama321b = 'mlx-community/Llama-3.2-1B-Instruct-4bit';

8 + static const String llama323b = 'mlx-community/Llama-3.2-3B-Instruct-4bit';

9 + static const String smolLm2135m = 'mlx-community/SmolLM2-135M-Instruct';

10 + static const String qwen2505b = 'mlx-community/Qwen2.5-0.5B-Instruct-4bit';

11 + static const String qwen2515b = 'mlx-community/Qwen2.5-1.5B-Instruct-4bit';

12 + static const String qwen253b = 'mlx-community/Qwen2.5-3B-Instruct-4bit';

13 +

14 + // Speech-To-Text (STT) Models

15 + static const String parakeetTdt06b = 'mlx-community/parakeet-tdt-0.6b-v2';

16 + static const String whisperTiny = 'mlx-community/whisper-tiny-mlx';

17 + static const String whisperBase = 'mlx-community/whisper-base-mlx';

18 + static const String whisperSmall = 'mlx-community/whisper-small-mlx';

19 +

20 + // Text-To-Speech (TTS) Models

21 + static const String pocketTts = 'mlx-community/pocket-tts';

22 +

23 + // Default PocketTTS voice IDs

24 + static const List<String> defaultVoices = [

25 + 'alba',

26 + 'marius',

27 + 'javert',

28 + 'jean',

29 + ];

30 +}

31 +

32 +/// Configuration options for MLX speech operations.

33 +class AppleMlxConfig {

34 + /// Path or Hugging Face repo ID for STT model.

35 + final String? sttModelId;

36 +

37 + /// Path or Hugging Face repo ID for TTS model.

38 + final String? ttsModelId;

39 +

40 + /// Default voice ID to use for TTS.

41 + final String defaultVoice;

42 +

43 + /// Maximum samples allowed in audio buffer for STT.

44 + final int? maximumInputSamples;

45 +

46 + const AppleMlxConfig({

47 + this.sttModelId = AppleMlxModelIds.parakeetTdt06b,

48 + this.ttsModelId = AppleMlxModelIds.pocketTts,

49 + this.defaultVoice = 'alba',

50 + this.maximumInputSamples,

51 + });

52 +}

packages/apple_mlx/lib/src/mlx_speech_manager.dart new

+107

@@ -0,0 +1,107 @@

1 +import 'dart:async';

2 +import 'package:mlx_audio/mlx_audio.dart' as mlx_native;

3 +import 'package:speech_mlx/speech_mlx.dart';

4 +import 'mlx_models.dart';

5 +

6 +/// Manager for orchestrating MLX Audio STT, TTS, and Turn Completion.

7 +class AppleMlxSpeechManager {

8 + final AppleMlxConfig config;

9 + MlxSpeechProvider? _provider;

10 + MlxSmartTurnScorer? _turnScorer;

11 + bool _isInitialized = false;

12 +

13 + AppleMlxSpeechManager({this.config = const AppleMlxConfig()});

14 +

15 + bool get isInitialized => _isInitialized;

16 +

17 + /// Initializes the provider with configured STT and TTS isolate workers.

18 + Future<void> initialize() async {

19 + if (_isInitialized) return;

20 +

21 + final sttWorker = config.sttModelId != null

22 + ? MlxIsolateBatchWorker(modelPath: config.sttModelId!)

23 + : null;

24 +

25 + final ttsWorker = config.ttsModelId != null

26 + ? MlxIsolateTtsWorker(modelPath: config.ttsModelId!)

27 + : null;

28 +

29 + _provider = MlxSpeechProvider(

30 + worker: sttWorker,

31 + modelId: config.sttModelId ?? 'default',

32 + ttsWorker: ttsWorker,

33 + ttsModelId: config.ttsModelId ?? 'pocket-tts',

34 + voiceIds: AppleMlxModelIds.defaultVoices,

35 + );

36 +

37 + _isInitialized = true;

38 + }

39 +

40 + /// Transcribes a WAV audio file directly using MLX Audio.

41 + Future<String> transcribeWavFile(String wavFilePath) async {

42 + final audio = await mlx_native.loadAudioMono16k(wavFilePath);

43 + final modelId = config.sttModelId ?? AppleMlxModelIds.parakeetTdt06b;

44 + final stt = await mlx_native.STT.loadModel(modelId);

45 + try {

46 + final output = stt.generate(audio);

47 + return output.text;

48 + } finally {

49 + stt.close();

50 + }

51 + }

52 +

53 + /// Synthesizes text to a WAV audio file using PocketTTS on MLX.

54 + Future<mlx_native.TtsOutput> synthesizeToWav({

55 + required String text,

56 + required String outputPath,

57 + String? voice,

58 + void Function(mlx_native.TtsAudioChunk chunk)? onAudioChunk,

59 + }) async {

60 + final modelId = config.ttsModelId ?? AppleMlxModelIds.pocketTts;

61 + final tts = await mlx_native.TTS.loadModel(modelId);

62 + try {

63 + final selectedVoice = voice ?? config.defaultVoice;

64 + final out = tts.generate(

65 + text,

66 + voice: selectedVoice,

67 + onAudioChunk: (chunk) {

68 + onAudioChunk?.call(chunk);

69 + },

70 + );

71 + mlx_native.writeWavFloat32(outputPath, out.samples, out.sampleRate);

72 + return out;

73 + } finally {

74 + tts.close();

75 + }

76 + }

77 +

78 + /// Initializes SmartTurn endpoint / turn-completion detector.

79 + Future<void> initializeTurnScorer({String? modelDirectory}) async {

80 + if (_turnScorer != null) return;

81 + _turnScorer = MlxSmartTurnScorer(

82 + worker: MlxIsolateTurnWorker(modelDirectory: modelDirectory),

83 + );

84 + }

85 +

86 + /// Evaluates whether user speech is complete using SmartTurn classifier.

87 + Future<mlx_native.SmartTurnEndpointOutput> scoreTurn(

88 + String pauseWavPath,

89 + ) async {

90 + final turn = await mlx_native.SmartTurn.load();

91 + try {

92 + final window = await mlx_native.loadAudioMono16k(pauseWavPath);

93 + return turn.predictEndpoint(window);

94 + } finally {

95 + turn.close();

96 + }

97 + }

98 +

99 + /// Disposes and cleans up MLX workers and isolates.

100 + Future<void> dispose() async {

101 + await _provider?.close();

102 + await _turnScorer?.close();

103 + _provider = null;

104 + _turnScorer = null;

105 + _isInitialized = false;

106 + }

107 +}

packages/apple_mlx/pubspec.lock new

+301

@@ -0,0 +1,301 @@

1 +# Generated by pub

2 +# See https://dart.dev/tools/pub/glossary#lockfile

3 +packages:

4 + args:

5 + dependency: transitive

6 + description:

7 + name: args

8 + sha256: d0481093c50b1da8910eb0bb301626d4d8eb7284aa739614d2b394ee09e3ea04

9 + url: "https://pub.dev"

10 + source: hosted

11 + version: "2.7.0"

12 + async:

13 + dependency: transitive

14 + description:

15 + name: async

16 + sha256: e2eb0491ba5ddb6177742d2da23904574082139b07c1e33b8503b9f46f3e1a37

17 + url: "https://pub.dev"

18 + source: hosted

19 + version: "2.13.1"

20 + audio_core:

21 + dependency: "direct main"

22 + description:

23 + name: audio_core

24 + sha256: "6996bd8f5df465a6fa8d08736487bd53f8b78d4ffee411fcfa4dbfdc2e2437ad"

25 + url: "https://pub.dev"

26 + source: hosted

27 + version: "0.1.0"

28 + boolean_selector:

29 + dependency: transitive

30 + description:

31 + name: boolean_selector

32 + sha256: "8aab1771e1243a5063b8b0ff68042d67334e3feab9e95b9490f9a6ebf73b42ea"

33 + url: "https://pub.dev"

34 + source: hosted

35 + version: "2.1.2"

36 + characters:

37 + dependency: transitive

38 + description:

39 + name: characters

40 + sha256: faf38497bda5ead2a8c7615f4f7939df04333478bf32e4173fcb06d428b5716b

41 + url: "https://pub.dev"

42 + source: hosted

43 + version: "1.4.1"

44 + clock:

45 + dependency: transitive

46 + description:

47 + name: clock

48 + sha256: fddb70d9b5277016c77a80201021d40a2247104d9f4aa7bab7157b7e3f05b84b

49 + url: "https://pub.dev"

50 + source: hosted

51 + version: "1.1.2"

52 + collection:

53 + dependency: transitive

54 + description:

55 + name: collection

56 + sha256: "2f5709ae4d3d59dd8f7cd309b4e023046b57d8a6c82130785d2b0e5868084e76"

57 + url: "https://pub.dev"

58 + source: hosted

59 + version: "1.19.1"

60 + crypto:

61 + dependency: transitive

62 + description:

63 + name: crypto

64 + sha256: c8ea0233063ba03258fbcf2ca4d6dadfefe14f02fab57702265467a19f27fadf

65 + url: "https://pub.dev"

66 + source: hosted

67 + version: "3.0.7"

68 + fake_async:

69 + dependency: transitive

70 + description:

71 + name: fake_async

72 + sha256: "5368f224a74523e8d2e7399ea1638b37aecfca824a3cc4dfdf77bf1fa905ac44"

73 + url: "https://pub.dev"

74 + source: hosted

75 + version: "1.3.3"

76 + ffi:

77 + dependency: transitive

78 + description:

79 + name: ffi

80 + sha256: "6d7fd89431262d8f3125e81b50d3847a091d846eafcd4fdb88dd06f36d705a45"

81 + url: "https://pub.dev"

82 + source: hosted

83 + version: "2.2.0"

84 + flutter:

85 + dependency: "direct main"

86 + description: flutter

87 + source: sdk

88 + version: "0.0.0"

89 + flutter_test:

90 + dependency: "direct dev"

91 + description: flutter

92 + source: sdk

93 + version: "0.0.0"

94 + http:

95 + dependency: "direct main"

96 + description:

97 + name: http

98 + sha256: "87721a4a50b19c7f1d49001e51409bddc46303966ce89a65af4f4e6004896412"

99 + url: "https://pub.dev"

100 + source: hosted

101 + version: "1.6.0"

102 + http_parser:

103 + dependency: transitive

104 + description:

105 + name: http_parser

106 + sha256: "178d74305e7866013777bab2c3d8726205dc5a4dd935297175b19a23a2e66571"

107 + url: "https://pub.dev"

108 + source: hosted

109 + version: "4.1.2"

110 + leak_tracker:

111 + dependency: transitive

112 + description:

113 + name: leak_tracker

114 + sha256: "33e2e26bdd85a0112ec15400c8cbffea70d0f9c3407491f672a2fad47915e2de"

115 + url: "https://pub.dev"

116 + source: hosted

117 + version: "11.0.2"

118 + leak_tracker_flutter_testing:

119 + dependency: transitive

120 + description:

121 + name: leak_tracker_flutter_testing

122 + sha256: "1dbc140bb5a23c75ea9c4811222756104fbcd1a27173f0c34ca01e16bea473c1"

123 + url: "https://pub.dev"

124 + source: hosted

125 + version: "3.0.10"

126 + leak_tracker_testing:

127 + dependency: transitive

128 + description:

129 + name: leak_tracker_testing

130 + sha256: "8d5a2d49f4a66b49744b23b018848400d23e54caf9463f4eb20df3eb8acb2eb1"

131 + url: "https://pub.dev"

132 + source: hosted

133 + version: "3.0.2"

134 + lib_mlx:

135 + dependency: "direct main"

136 + description:

137 + name: lib_mlx

138 + sha256: dcb32359d834425cabb820de7de051a8f85d536e96218abb1d68d5973cf7044d

139 + url: "https://pub.dev"

140 + source: hosted

141 + version: "0.1.1"

142 + matcher:

143 + dependency: transitive

144 + description:

145 + name: matcher

146 + sha256: dc0b7dc7651697ea4ff3e69ef44b0407ea32c487a39fff6a4004fa585e901861

147 + url: "https://pub.dev"

148 + source: hosted

149 + version: "0.12.19"

150 + material_color_utilities:

151 + dependency: transitive

152 + description:

153 + name: material_color_utilities

154 + sha256: "9c337007e82b1889149c82ed242ed1cb24a66044e30979c44912381e9be4c48b"

155 + url: "https://pub.dev"

156 + source: hosted

157 + version: "0.13.0"

158 + meta:

159 + dependency: transitive

160 + description:

161 + name: meta

162 + sha256: "1741988757a65eb6b36abe716829688cf01910bbf91c34354ff7ec1c3de2b349"

163 + url: "https://pub.dev"

164 + source: hosted

165 + version: "1.18.0"

166 + mlx:

167 + dependency: transitive

168 + description:

169 + name: mlx

170 + sha256: d735465e4a8f55dc209d7e53aaf3039b9063101aa3cced0da2843b64e884a721

171 + url: "https://pub.dev"

172 + source: hosted

173 + version: "0.2.0"

174 + mlx_audio:

175 + dependency: "direct main"

176 + description:

177 + name: mlx_audio

178 + sha256: "5a94ffaea1bce9e75f3757722e1a92dd8514a3c8bb139377a07b4e7f6c304874"

179 + url: "https://pub.dev"

180 + source: hosted

181 + version: "0.2.2"

182 + path:

183 + dependency: transitive

184 + description:

185 + name: path

186 + sha256: "75cca69d1490965be98c73ceaea117e8a04dd21217b37b292c9ddbec0d955bc5"

187 + url: "https://pub.dev"

188 + source: hosted

189 + version: "1.9.1"

190 + sky_engine:

191 + dependency: transitive

192 + description: flutter

193 + source: sdk

194 + version: "0.0.0"

195 + source_span:

196 + dependency: transitive

197 + description:

198 + name: source_span

199 + sha256: "56a02f1f4cd1a2d96303c0144c93bd6d909eea6bee6bf5a0e0b685edbd4c47ab"

200 + url: "https://pub.dev"

201 + source: hosted

202 + version: "1.10.2"

203 + speech_core:

204 + dependency: "direct main"

205 + description:

206 + name: speech_core

207 + sha256: bc0cd35c890528638ba4a55792e05a8b5afe23bfc4b2ba30873ebb3813c97bad

208 + url: "https://pub.dev"

209 + source: hosted

210 + version: "0.2.0"

211 + speech_mlx:

212 + dependency: "direct main"

213 + description:

214 + name: speech_mlx

215 + sha256: "35aae05aa0e5b94695fd0038cb2dcada83e4c310e0268e4ef29a3ece7973bd25"

216 + url: "https://pub.dev"

217 + source: hosted

218 + version: "0.2.0"

219 + stack_trace:

220 + dependency: transitive

221 + description:

222 + name: stack_trace

223 + sha256: "8b27215b45d22309b5cddda1aa2b19bdfec9df0e765f2de506401c071d38d1b1"

224 + url: "https://pub.dev"

225 + source: hosted

226 + version: "1.12.1"

227 + stream_channel:

228 + dependency: transitive

229 + description:

230 + name: stream_channel

231 + sha256: "969e04c80b8bcdf826f8f16579c7b14d780458bd97f56d107d3950fdbeef059d"

232 + url: "https://pub.dev"

233 + source: hosted

234 + version: "2.1.4"

235 + string_scanner:

236 + dependency: transitive

237 + description:

238 + name: string_scanner

239 + sha256: "921cd31725b72fe181906c6a94d987c78e3b98c2e205b397ea399d4054872b43"

240 + url: "https://pub.dev"

241 + source: hosted

242 + version: "1.4.1"

243 + term_glyph:

244 + dependency: transitive

245 + description:

246 + name: term_glyph

247 + sha256: "7f554798625ea768a7518313e58f83891c7f5024f88e46e7182a4558850a4b8e"

248 + url: "https://pub.dev"

249 + source: hosted

250 + version: "1.2.2"

251 + test_api:

252 + dependency: transitive

253 + description:

254 + name: test_api

255 + sha256: "949a932224383300f01be9221c39180316445ecb8e7547f70a41a35bf421fb9e"

256 + url: "https://pub.dev"

257 + source: hosted

258 + version: "0.7.11"

259 + typed_data:

260 + dependency: transitive

261 + description:

262 + name: typed_data

263 + sha256: f9049c039ebfeb4cf7a7104a675823cd72dba8297f264b6637062516699fa006

264 + url: "https://pub.dev"

265 + source: hosted

266 + version: "1.4.0"

267 + vector_math:

268 + dependency: transitive

269 + description:

270 + name: vector_math

271 + sha256: d530bd74fea330e6e364cda7a85019c434070188383e1cd8d9777ee586914c5b

272 + url: "https://pub.dev"

273 + source: hosted

274 + version: "2.2.0"

275 + vm_service:

276 + dependency: transitive

277 + description:

278 + name: vm_service

279 + sha256: "5f37239c4851efcef929cea7824e76df7f2f0970aef85d66bbc430afa40e72f0"

280 + url: "https://pub.dev"

281 + source: hosted

282 + version: "15.3.0"

283 + wav:

284 + dependency: transitive

285 + description:

286 + name: wav

287 + sha256: "2c932c74a531befedcb1d44bb24dd95fea8c801590d560c73fa94f35c5ef6551"

288 + url: "https://pub.dev"

289 + source: hosted

290 + version: "1.5.0"

291 + web:

292 + dependency: transitive

293 + description:

294 + name: web

295 + sha256: "868d88a33d8a87b18ffc05f9f030ba328ffefba92d6c127917a2ba740f9cfe4a"

296 + url: "https://pub.dev"

297 + source: hosted

298 + version: "1.1.1"

299 +sdks:

300 + dart: ">=3.12.0 <4.0.0"

301 + flutter: ">=3.44.0"

packages/apple_mlx/pubspec.yaml new

+22

@@ -0,0 +1,22 @@

1 +name: apple_mlx

2 +description: Local package bridging Apple MLX models (LLM text generation, STT, TTS, turn detection) for LocalMind.

3 +version: 0.1.0

4 +publish_to: none

5 +

6 +environment:

7 + sdk: ">=3.12.0 <4.0.0"

8 + flutter: ">=3.44.0"

9 +

10 +dependencies:

11 + flutter:

12 + sdk: flutter

13 + lib_mlx: ^0.1.1

14 + speech_mlx: ^0.2.0

15 + mlx_audio: ^0.2.2

16 + speech_core: ^0.2.0

17 + audio_core: ^0.1.0

18 + http: ^1.2.0

19 +

20 +dev_dependencies:

21 + flutter_test:

22 + sdk: flutter

packages/apple_mlx/test/apple_mlx_test.dart new

+44

@@ -0,0 +1,44 @@

1 +import 'package:flutter_test/flutter_test.dart';

2 +import 'package:apple_mlx/apple_mlx.dart';

3 +

4 +void main() {

5 + group('AppleMlxSpeechManager', () {

6 + test('initializes with default configuration', () {

7 + final manager = AppleMlxSpeechManager();

8 + expect(manager.isInitialized, isFalse);

9 + expect(manager.config.defaultVoice, equals('alba'));

10 + expect(

11 + manager.config.sttModelId,

12 + equals(AppleMlxModelIds.parakeetTdt06b),

13 + );

14 + expect(manager.config.ttsModelId, equals(AppleMlxModelIds.pocketTts));

15 + });

16 +

17 + test('supports custom model configuration', () {

18 + final config = AppleMlxConfig(

19 + sttModelId: AppleMlxModelIds.whisperSmall,

20 + ttsModelId: AppleMlxModelIds.pocketTts,

21 + defaultVoice: 'marius',

22 + );

23 + final manager = AppleMlxSpeechManager(config: config);

24 + expect(manager.config.defaultVoice, equals('marius'));

25 + expect(manager.config.sttModelId, equals(AppleMlxModelIds.whisperSmall));

26 + });

27 + });

28 +

29 + group('AppleMlxLlmManager', () {

30 + test('instantiates with uninitialized state', () {

31 + final llm = AppleMlxLlmManager();

32 + expect(llm.isModelLoaded, isFalse);

33 + expect(llm.isServerRunning, isFalse);

34 + expect(llm.serverInfo, isNull);

35 + expect(llm.modelHandle, isNull);

36 + });

37 +

38 + test('exposes popular MLX model identifiers', () {

39 + expect(AppleMlxModelIds.gemma22b, isNotEmpty);

40 + expect(AppleMlxModelIds.llama321b, isNotEmpty);

41 + expect(AppleMlxModelIds.qwen2505b, isNotEmpty);

42 + });

43 + });

44 +}

pubspec.lock

+63

@@ -25,6 +25,13 @@ packages:

25 url: "https://pub.dev"

26 source: hosted

27 version: "2.2.0"

28 + apple_mlx:

29 + dependency: "direct main"

30 + description:

31 + path: "packages/apple_mlx"

32 + relative: true

33 + source: path

34 + version: "0.1.0"

35 archive:

36 dependency: "direct main"

37 description:

@@ -49,6 +56,14 @@ packages:

56 url: "https://pub.dev"

57 source: hosted

58 version: "2.13.1"

59 + audio_core:

60 + dependency: transitive

61 + description:

62 + name: audio_core

63 + sha256: "6996bd8f5df465a6fa8d08736487bd53f8b78d4ffee411fcfa4dbfdc2e2437ad"

64 + url: "https://pub.dev"

65 + source: hosted

66 + version: "0.1.0"

67 audio_service:

68 dependency: transitive

69 description:

@@ -1093,6 +1108,14 @@ packages:

1108 url: "https://pub.dev"

1109 source: hosted

1110 version: "3.0.2"

1111 + lib_mlx:

1112 + dependency: transitive

1113 + description:

1114 + name: lib_mlx

1115 + sha256: dcb32359d834425cabb820de7de051a8f85d536e96218abb1d68d5973cf7044d

1116 + url: "https://pub.dev"

1117 + source: hosted

1118 + version: "0.1.1"

1119 lints:

1120 dependency: transitive

1121 description:

@@ -1181,6 +1204,22 @@ packages:

1204 url: "https://pub.dev"

1205 source: hosted

1206 version: "2.0.0"

1207 + mlx:

1208 + dependency: transitive

1209 + description:

1210 + name: mlx

1211 + sha256: d735465e4a8f55dc209d7e53aaf3039b9063101aa3cced0da2843b64e884a721

1212 + url: "https://pub.dev"

1213 + source: hosted

1214 + version: "0.2.0"

1215 + mlx_audio:

1216 + dependency: transitive

1217 + description:

1218 + name: mlx_audio

1219 + sha256: "5a94ffaea1bce9e75f3757722e1a92dd8514a3c8bb139377a07b4e7f6c304874"

1220 + url: "https://pub.dev"

1221 + source: hosted

1222 + version: "0.2.2"

1223 mutex:

1224 dependency: transitive

1225 description:

@@ -1762,6 +1801,22 @@ packages:

1801 url: "https://pub.dev"

1802 source: hosted

1803 version: "1.10.2"

1804 + speech_core:

1805 + dependency: transitive

1806 + description:

1807 + name: speech_core

1808 + sha256: bc0cd35c890528638ba4a55792e05a8b5afe23bfc4b2ba30873ebb3813c97bad

1809 + url: "https://pub.dev"

1810 + source: hosted

1811 + version: "0.2.0"

1812 + speech_mlx:

1813 + dependency: transitive

1814 + description:

1815 + name: speech_mlx

1816 + sha256: "35aae05aa0e5b94695fd0038cb2dcada83e4c310e0268e4ef29a3ece7973bd25"

1817 + url: "https://pub.dev"

1818 + source: hosted

1819 + version: "0.2.0"

1820 speech_to_text:

1821 dependency: "direct main"

1822 description:

@@ -2099,6 +2154,14 @@ packages:

2154 url: "https://pub.dev"

2155 source: hosted

2156 version: "1.2.1"

2157 + wav:

2158 + dependency: transitive

2159 + description:

2160 + name: wav

2161 + sha256: "2c932c74a531befedcb1d44bb24dd95fea8c801590d560c73fa94f35c5ef6551"

2162 + url: "https://pub.dev"

2163 + source: hosted

2164 + version: "1.5.0"

2165 web:

2166 dependency: transitive

2167 description:

pubspec.yaml

@@ -89,6 +89,8 @@ dependencies:

89 auto_size_text_plus: ^3.0.2

90 device_calendar_plus: ^0.8.0

91 home_widget: ^0.9.3

92 + apple_mlx:

93 + path: packages/apple_mlx

95 dev_dependencies:

96 flutter_test:

test/features/os_widget/os_widget_snapshot_provider_test.dart

+67 -60

@@ -28,26 +28,29 @@ class _StubServersNotifier extends ServersNotifier {

29 void main() {

30 group('osWidgetSnapshotProvider', () {

31 - test('returns unconfigured snapshot when no default model is set', () async {

32 - SharedPreferences.setMockInitialValues({});

33 - final prefs = await SharedPreferences.getInstance();

34 -

35 - final container = ProviderContainer(

36 - overrides: [

37 - sharedPreferencesProvider.overrideWithValue(prefs),

38 - settingsProvider.overrideWith(_TestSettingsNotifier.new),

39 - serversProvider.overrideWith(_StubServersNotifier.new),

40 - ],

41 - );

42 - addTearDown(container.dispose);

43 -

44 - // Force settings to build

45 - container.read(settingsProvider);

46 -

47 - final snapshot = container.read(osWidgetSnapshotProvider);

48 - expect(snapshot.isConfigured, isFalse);

49 - expect(snapshot.modelName, isNull);

50 - });

31 + test(

32 + 'returns unconfigured snapshot when no default model is set',

33 + () async {

34 + SharedPreferences.setMockInitialValues({});

35 + final prefs = await SharedPreferences.getInstance();

36 +

37 + final container = ProviderContainer(

38 + overrides: [

39 + sharedPreferencesProvider.overrideWithValue(prefs),

40 + settingsProvider.overrideWith(_TestSettingsNotifier.new),

41 + serversProvider.overrideWith(_StubServersNotifier.new),

42 + ],

43 + );

44 + addTearDown(container.dispose);

45 +

46 + // Force settings to build

47 + container.read(settingsProvider);

48 +

49 + final snapshot = container.read(osWidgetSnapshotProvider);

50 + expect(snapshot.isConfigured, isFalse);

51 + expect(snapshot.modelName, isNull);

52 + },

53 + );

55 test('returns configured snapshot with model display name', () async {

56 SharedPreferences.setMockInitialValues({});

@@ -81,7 +84,9 @@ void main() {

84 () => _TestSettingsNotifier(customSettings: settings),

85 ),

86 serversProvider.overrideWith(() => _StubServersNotifier([server])),

84 - availableModelsProvider(server.id).overrideWith((ref) async => [model]),

87 + availableModelsProvider(

88 + server.id,

89 + ).overrideWith((ref) async => [model]),

90 ],

91 );

92 addTearDown(container.dispose);

@@ -94,43 +99,45 @@ void main() {

99 expect(snapshot.modelName, 'Llama 3 8B');

100 });

101

97 - test('falls back to model id when model is missing from available list',

98 - () async {

99 - SharedPreferences.setMockInitialValues({});

100 - final prefs = await SharedPreferences.getInstance();

101 -

102 - final server = Server(

103 - id: 'srv-1',

104 - name: 'Test Server',

105 - host: 'http://localhost',

106 - port: 11434,

107 - type: ServerType.ollama,

108 - createdAt: DateTime.now(),

109 - lastConnectedAt: DateTime.now(),

110 - );

111 -

112 - final settings = AppSettings(

113 - defaultModelId: 'missing-model',

114 - defaultModelServerId: server.id,

115 - );

116 -

117 - final container = ProviderContainer(

118 - overrides: [

119 - sharedPreferencesProvider.overrideWithValue(prefs),

120 - settingsProvider.overrideWith(

121 - () => _TestSettingsNotifier(customSettings: settings),

122 - ),

123 - serversProvider.overrideWith(() => _StubServersNotifier([server])),

124 - availableModelsProvider(server.id).overrideWith((ref) async => []),

125 - ],

126 - );

127 - addTearDown(container.dispose);

128 -

129 - await container.read(availableModelsProvider(server.id).future);

130 -

131 - final snapshot = container.read(osWidgetSnapshotProvider);

132 - expect(snapshot.isConfigured, isTrue);

133 - expect(snapshot.modelName, 'missing-model');

134 - });

102 + test(

103 + 'falls back to model id when model is missing from available list',

104 + () async {

105 + SharedPreferences.setMockInitialValues({});

106 + final prefs = await SharedPreferences.getInstance();

107 +

108 + final server = Server(

109 + id: 'srv-1',

110 + name: 'Test Server',

111 + host: 'http://localhost',

112 + port: 11434,

113 + type: ServerType.ollama,

114 + createdAt: DateTime.now(),

115 + lastConnectedAt: DateTime.now(),

116 + );

117 +

118 + final settings = AppSettings(

119 + defaultModelId: 'missing-model',

120 + defaultModelServerId: server.id,

121 + );

122 +

123 + final container = ProviderContainer(

124 + overrides: [

125 + sharedPreferencesProvider.overrideWithValue(prefs),

126 + settingsProvider.overrideWith(

127 + () => _TestSettingsNotifier(customSettings: settings),

128 + ),

129 + serversProvider.overrideWith(() => _StubServersNotifier([server])),

130 + availableModelsProvider(server.id).overrideWith((ref) async => []),

131 + ],

132 + );

133 + addTearDown(container.dispose);

134 +

135 + await container.read(availableModelsProvider(server.id).future);

136 +

137 + final snapshot = container.read(osWidgetSnapshotProvider);

138 + expect(snapshot.isConfigured, isTrue);

139 + expect(snapshot.modelName, 'missing-model');

140 + },

141 + );

142 });

136 -}

\ No newline at end of file

143 +}