webml-kit 0.4.0 → 0.4.1

package/dist/model-client.js

@@ -1,318 +0,0 @@
-/**
- * Main-thread client for communicating with the ML worker.
- *
- * Wraps the Web Worker's postMessage protocol with a promise-based API,
- * AsyncIterable token streaming, and typed events.
- *
- * ```ts
- * import { ModelClient } from 'webml-kit';
- *
- * const client = new ModelClient();
- *
- * // Check device capabilities
- * const device = await client.detect();
- * console.log(device.backend); // 'webgpu'
- *
- * // Load a model
- * await client.load({
- *   task: 'text-generation',
- *   modelId: 'onnx-community/Bonsai-1.7B-ONNX',
- *   dtype: 'q4',
- *   onProgress: ({ percent }) => console.log(`${percent}%`),
- * });
- *
- * // Stream text generation
- * for await (const { token, tps } of client.stream('Hello!')) {
- *   process.stdout.write(token);
- * }
- *
- * // One-shot inference
- * const labels = await client.run('image-classification', imageBlob);
- * ```
- */
-import { TokenStream, collectStream } from './streaming.js';
-import { GPURecovery } from './gpu-recovery.js';
-// ─── Client ───
-export class ModelClient {
-    worker = null;
-    workerUrl;
-    listeners = new Map();
-    pendingRequests = new Map();
-    requestCounter = 0;
-    deviceInfo = null;
-    loadedModels = new Set();
-    progressCallback = null;
-    gpuRecovery;
-    /**
-     * Create a new ModelClient.
-     *
-     * @param workerUrl - URL to the model-worker.js file. If omitted,
-     *   creates a Blob URL from the bundled worker (requires bundler support).
-     */
-    constructor(workerUrl) {
-        this.workerUrl = workerUrl ?? null;
-        this.gpuRecovery = new GPURecovery();
-        this.gpuRecovery.on('lost', ({ reason }) => {
-            this.emit('device-lost', { reason });
-        });
-        this.gpuRecovery.on('recovered', () => {
-            this.emit('device-recovered', {});
-        });
-    }
-    // ─── Worker Lifecycle ───
-    getWorker() {
-        if (!this.worker) {
-            const url = this.workerUrl
-                ?? new URL('./model-worker.js', import.meta.url);
-            this.worker = new Worker(url, { type: 'module' });
-            this.worker.addEventListener('message', this.handleMessage.bind(this));
-            this.worker.addEventListener('error', (e) => {
-                this.emit('error', { message: e.message });
-            });
-        }
-        return this.worker;
-    }
-    send(cmd) {
-        this.getWorker().postMessage(cmd);
-    }
-    nextId() {
-        return `req_${++this.requestCounter}_${Date.now()}`;
-    }
-    // ─── Message Handler ───
-    handleMessage(e) {
-        const msg = e.data;
-        switch (msg.type) {
-            case 'device-info': {
-                this.deviceInfo = msg.data;
-                const pending = this.pendingRequests.get('detect');
-                if (pending) {
-                    pending.resolve(msg.data);
-                    this.pendingRequests.delete('detect');
-                }
-                break;
-            }
-            case 'progress': {
-                this.progressCallback?.(msg.data);
-                this.emit('progress', msg.data);
-                break;
-            }
-            case 'ready': {
-                this.loadedModels.add(msg.modelKey);
-                this.emit('ready', { modelKey: msg.modelKey });
-                const pending = this.pendingRequests.get('load');
-                if (pending) {
-                    pending.resolve(undefined);
-                    this.pendingRequests.delete('load');
-                }
-                break;
-            }
-            case 'token': {
-                const pending = this.pendingRequests.get(msg.id);
-                if (pending?.stream) {
-                    pending.stream.push(msg.data);
-                }
-                break;
-            }
-            case 'result': {
-                const pending = this.pendingRequests.get(msg.id);
-                if (pending) {
-                    if (pending.stream) {
-                        pending.stream.end();
-                    }
-                    pending.resolve(msg.data);
-                    this.pendingRequests.delete(msg.id);
-                }
-                break;
-            }
-            case 'error': {
-                const pending = this.pendingRequests.get(msg.id);
-                if (pending) {
-                    if (pending.stream) {
-                        pending.stream.abort(new Error(msg.data));
-                    }
-                    pending.reject(new Error(msg.data));
-                    this.pendingRequests.delete(msg.id);
-                }
-                this.emit('error', { message: msg.data, id: msg.id });
-                break;
-            }
-            case 'device-lost': {
-                this.emit('device-lost', { reason: msg.reason });
-                break;
-            }
-            case 'device-recovered': {
-                this.emit('device-recovered', {});
-                break;
-            }
-        }
-    }
-    // ─── Events ───
-    /** Register an event listener. */
-    on(event, listener) {
-        if (!this.listeners.has(event)) {
-            this.listeners.set(event, new Set());
-        }
-        this.listeners.get(event).add(listener);
-        return this;
-    }
-    /** Remove an event listener. */
-    off(event, listener) {
-        this.listeners.get(event)?.delete(listener);
-        return this;
-    }
-    emit(event, data) {
-        this.listeners.get(event)?.forEach(fn => fn(data));
-    }
-    // ─── Public API ───
-    /**
-     * Detect the best available compute backend.
-     *
-     * ```ts
-     * const info = await client.detect();
-     * console.log(info.backend);         // 'webgpu'
-     * console.log(info.gpu?.vendor);      // 'apple'
-     * console.log(info.recommendedDtype); // 'q4'
-     * ```
-     */
-    async detect() {
-        if (this.deviceInfo)
-            return this.deviceInfo;
-        return new Promise((resolve, reject) => {
-            this.pendingRequests.set('detect', { resolve: resolve, reject });
-            this.send({ type: 'check' });
-        });
-    }
-    /**
-     * Load a model pipeline.
-     *
-     * ```ts
-     * await client.load({
-     *   task: 'text-generation',
-     *   modelId: 'onnx-community/Bonsai-1.7B-ONNX',
-     *   dtype: 'q4',
-     *   onProgress: ({ percent }) => updateUI(percent),
-     * });
-     * ```
-     */
-    async load(options) {
-        this.progressCallback = options.onProgress ?? null;
-        const config = {
-            task: options.task,
-            modelId: options.modelId,
-            dtype: options.dtype,
-            device: options.device,
-            revision: options.revision,
-        };
-        return new Promise((resolve, reject) => {
-            this.pendingRequests.set('load', { resolve: resolve, reject });
-            this.send({ type: 'load', config });
-        });
-    }
-    /**
-     * Run one-shot inference for any pipeline task.
-     *
-     * ```ts
-     * // Image classification
-     * const labels = await client.run('image-classification', imageUrl);
-     *
-     * // Speech recognition
-     * const { text } = await client.run('automatic-speech-recognition', audioBlob);
-     *
-     * // Embeddings
-     * const vectors = await client.run('feature-extraction', 'Hello world');
-     * ```
-     */
-    async run(task, input, options) {
-        const id = this.nextId();
-        return new Promise((resolve, reject) => {
-            this.pendingRequests.set(id, { resolve: resolve, reject });
-            this.send({ type: 'run', id, task, input, options });
-        });
-    }
-    /**
-     * Generate text with streaming tokens.
-     *
-     * Returns an `AsyncIterable<TokenEvent>` that yields tokens as they're generated.
-     *
-     * ```ts
-     * for await (const { token, tps } of client.stream('Tell me a joke')) {
-     *   process.stdout.write(token);
-     * }
-     * ```
-     */
-    stream(input, options) {
-        const id = this.nextId();
-        const stream = new TokenStream(options?.signal);
-        this.pendingRequests.set(id, {
-            resolve: () => { }, // Result comes through the stream
-            reject: (err) => stream.abort(err),
-            stream,
-        });
-        this.send({
-            type: 'run',
-            id,
-            task: 'text-generation',
-            input,
-            options: options,
-        });
-        return stream;
-    }
-    /**
-     * Generate text and wait for the complete result.
-     *
-     * ```ts
-     * const { text, tps, numTokens } = await client.generate('Hello!');
-     * ```
-     */
-    async generate(input, options) {
-        const tokenStream = this.stream(input, options);
-        const collected = await collectStream(tokenStream);
-        return {
-            ...collected,
-            totalTime: 0, // Will be filled from worker result
-        };
-    }
-    /**
-     * Interrupt an ongoing text generation.
-     */
-    interrupt() {
-        this.send({ type: 'interrupt' });
-    }
-    /**
-     * Reset the KV cache (start a new conversation).
-     */
-    reset() {
-        this.send({ type: 'reset' });
-    }
-    /**
-     * Dispose a loaded model and free memory.
-     *
-     * @param modelKey - Specific model key (task::modelId), or omit to dispose all.
-     */
-    dispose(modelKey) {
-        this.send({ type: 'dispose', modelKey });
-        if (modelKey) {
-            this.loadedModels.delete(modelKey);
-        }
-        else {
-            this.loadedModels.clear();
-        }
-    }
-    /**
-     * Check if a model is currently loaded.
-     */
-    isLoaded(task, modelId) {
-        return this.loadedModels.has(`${task}::${modelId}`);
-    }
-    /**
-     * Terminate the worker completely.
-     */
-    terminate() {
-        this.worker?.terminate();
-        this.worker = null;
-        this.loadedModels.clear();
-        this.pendingRequests.clear();
-        this.deviceInfo = null;
-    }
-}
-//# sourceMappingURL=model-client.js.map

package/dist/model-client.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"model-client.js","sourceRoot":"","sources":["../src/model-client.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AAeH,OAAO,EAAE,WAAW,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAC5D,OAAO,EAAE,WAAW,EAAE,MAAM,mBAAmB,CAAC;AAiBhD,iBAAiB;AAEjB,MAAM,OAAO,WAAW;IACd,MAAM,GAAkB,IAAI,CAAC;IAC7B,SAAS,CAAsB;IAC/B,SAAS,GAAG,IAAI,GAAG,EAAkC,CAAC;IACtD,eAAe,GAAG,IAAI,GAAG,EAO9B,CAAC;IACI,cAAc,GAAG,CAAC,CAAC;IACnB,UAAU,GAAsB,IAAI,CAAC;IACrC,YAAY,GAAG,IAAI,GAAG,EAAU,CAAC;IACjC,gBAAgB,GAA4B,IAAI,CAAC;IACjD,WAAW,CAAc;IAEjC;;;;;OAKG;IACH,YAAY,SAAwB;QAClC,IAAI,CAAC,SAAS,GAAG,SAAS,IAAI,IAAI,CAAC;QACnC,IAAI,CAAC,WAAW,GAAG,IAAI,WAAW,EAAE,CAAC;QAErC,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE;YACzC,IAAI,CAAC,IAAI,CAAC,aAAa,EAAE,EAAE,MAAM,EAAE,CAAC,CAAC;QACvC,CAAC,CAAC,CAAC;QACH,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC,WAAW,EAAE,GAAG,EAAE;YACpC,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,EAAE,CAAC,CAAC;QACpC,CAAC,CAAC,CAAC;IACL,CAAC;IAED,2BAA2B;IAEnB,SAAS;QACf,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC;YACjB,MAAM,GAAG,GAAG,IAAI,CAAC,SAAS;mBACrB,IAAI,GAAG,CAAC,mBAAmB,EAAE,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;YAEnD,IAAI,CAAC,MAAM,GAAG,IAAI,MAAM,CAAC,GAAG,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC,CAAC;YAElD,IAAI,CAAC,MAAM,CAAC,gBAAgB,CAAC,SAAS,EAAE,IAAI,CAAC,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC;YACvE,IAAI,CAAC,MAAM,CAAC,gBAAgB,CAAC,OAAO,EAAE,CAAC,CAAC,EAAE,EAAE;gBAC1C,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC;YAC7C,CAAC,CAAC,CAAC;QACL,CAAC;QACD,OAAO,IAAI,CAAC,MAAM,CAAC;IACrB,CAAC;IAEO,IAAI,CAAC,GAAkB;QAC7B,IAAI,CAAC,SAAS,EAAE,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC;IACpC,CAAC;IAEO,MAAM;QACZ,OAAO,OAAO,EAAE,IAAI,CAAC,cAAc,IAAI,IAAI,CAAC,GAAG,EAAE,EAAE,CAAC;IACtD,CAAC;IAED,0BAA0B;IAElB,aAAa,CAAC,CAA+B;QACnD,MAAM,GAAG,GAAG,CAAC,CAAC,IAAI,CAAC;QAEnB,QAAQ,GAAG,CAAC,IAAI,EAAE,CAAC;YACjB,KAAK,aAAa,CAAC,CAAC,CAAC;gBACnB,IAAI,CAAC,UAAU,GAAG,GAAG,CAAC,IAAI,CAAC;gBAC3B,MAAM,OAAO,GAAG,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;gBACnD,IAAI,OAAO,EAAE,CAAC;oBACZ,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;oBAC1B,IAAI,CAAC,eAAe,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;gBACxC,CAAC;gBACD,MAAM;YACR,CAAC;YAED,KAAK,UAAU,CAAC,CAAC,CAAC;gBAChB,IAAI,CAAC,gBAAgB,EAAE,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;gBAClC,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,GAAG,CAAC,IAAI,CAAC,CAAC;gBAChC,MAAM;YACR,CAAC;YAED,KAAK,OAAO,CAAC,CAAC,CAAC;gBACb,IAAI,CAAC,YAAY,CAAC,GAAG,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;gBACpC,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,EAAE,QAAQ,EAAE,GAAG,CAAC,QAAQ,EAAE,CAAC,CAAC;gBAC/C,MAAM,OAAO,GAAG,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;gBACjD,IAAI,OAAO,EAAE,CAAC;oBACZ,OAAO,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC;oBAC3B,IAAI,CAAC,eAAe,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;gBACtC,CAAC;gBACD,MAAM;YACR,CAAC;YAED,KAAK,OAAO,CAAC,CAAC,CAAC;gBACb,MAAM,OAAO,GAAG,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;gBACjD,IAAI,OAAO,EAAE,MAAM,EAAE,CAAC;oBACpB,OAAO,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;gBAChC,CAAC;gBACD,MAAM;YACR,CAAC;YAED,KAAK,QAAQ,CAAC,CAAC,CAAC;gBACd,MAAM,OAAO,GAAG,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;gBACjD,IAAI,OAAO,EAAE,CAAC;oBACZ,IAAI,OAAO,CAAC,MAAM,EAAE,CAAC;wBACnB,OAAO,CAAC,MAAM,CAAC,GAAG,EAAE,CAAC;oBACvB,CAAC;oBACD,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;oBAC1B,IAAI,CAAC,eAAe,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;gBACtC,CAAC;gBACD,MAAM;YACR,CAAC;YAED,KAAK,OAAO,CAAC,CAAC,CAAC;gBACb,MAAM,OAAO,GAAG,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;gBACjD,IAAI,OAAO,EAAE,CAAC;oBACZ,IAAI,OAAO,CAAC,MAAM,EAAE,CAAC;wBACnB,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;oBAC5C,CAAC;oBACD,OAAO,CAAC,MAAM,CAAC,IAAI,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;oBACpC,IAAI,CAAC,eAAe,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;gBACtC,CAAC;gBACD,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,EAAE,OAAO,EAAE,GAAG,CAAC,IAAI,EAAE,EAAE,EAAE,GAAG,CAAC,EAAE,EAAE,CAAC,CAAC;gBACtD,MAAM;YACR,CAAC;YAED,KAAK,aAAa,CAAC,CAAC,CAAC;gBACnB,IAAI,CAAC,IAAI,CAAC,aAAa,EAAE,EAAE,MAAM,EAAE,GAAG,CAAC,MAAM,EAAE,CAAC,CAAC;gBACjD,MAAM;YACR,CAAC;YAED,KAAK,kBAAkB,CAAC,CAAC,CAAC;gBACxB,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,EAAE,CAAC,CAAC;gBAClC,MAAM;YACR,CAAC;QACH,CAAC;IACH,CAAC;IAED,iBAAiB;IAEjB,kCAAkC;IAClC,EAAE,CAAC,KAAsB,EAAE,QAAkB;QAC3C,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC;YAC/B,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,GAAG,EAAE,CAAC,CAAC;QACvC,CAAC;QACD,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,KAAK,CAAE,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;QACzC,OAAO,IAAI,CAAC;IACd,CAAC;IAED,gCAAgC;IAChC,GAAG,CAAC,KAAsB,EAAE,QAAkB;QAC5C,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC,QAAQ,CAAC,CAAC;QAC5C,OAAO,IAAI,CAAC;IACd,CAAC;IAEO,IAAI,CAAC,KAAsB,EAAE,IAAa;QAChD,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,OAAO,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC;IACrD,CAAC;IAED,qBAAqB;IAErB;;;;;;;;;OASG;IACH,KAAK,CAAC,MAAM;QACV,IAAI,IAAI,CAAC,UAAU;YAAE,OAAO,IAAI,CAAC,UAAU,CAAC;QAE5C,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;YACrC,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,QAAQ,EAAE,EAAE,OAAO,EAAE,OAA+B,EAAE,MAAM,EAAE,CAAC,CAAC;YACzF,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;QAC/B,CAAC,CAAC,CAAC;IACL,CAAC;IAED;;;;;;;;;;;OAWG;IACH,KAAK,CAAC,IAAI,CAAC,OAAoB;QAC7B,IAAI,CAAC,gBAAgB,GAAG,OAAO,CAAC,UAAU,IAAI,IAAI,CAAC;QAEnD,MAAM,MAAM,GAAgB;YAC1B,IAAI,EAAE,OAAO,CAAC,IAAI;YAClB,OAAO,EAAE,OAAO,CAAC,OAAO;YACxB,KAAK,EAAE,OAAO,CAAC,KAAK;YACpB,MAAM,EAAE,OAAO,CAAC,MAAM;YACtB,QAAQ,EAAE,OAAO,CAAC,QAAQ;SAC3B,CAAC;QAEF,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;YACrC,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,MAAM,EAAE,EAAE,OAAO,EAAE,OAA+B,EAAE,MAAM,EAAE,CAAC,CAAC;YACvF,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC,CAAC;QACtC,CAAC,CAAC,CAAC;IACL,CAAC;IAED;;;;;;;;;;;;;OAaG;IACH,KAAK,CAAC,GAAG,CACP,IAAkB,EAClB,KAAc,EACd,OAAiC;QAEjC,MAAM,EAAE,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC;QAEzB,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;YACrC,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,EAAE,EAAE,EAAE,OAAO,EAAE,OAA+B,EAAE,MAAM,EAAE,CAAC,CAAC;YACnF,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,KAAK,EAAE,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE,OAAO,EAAE,CAAC,CAAC;QACvD,CAAC,CAAC,CAAC;IACL,CAAC;IAED;;;;;;;;;;OAUG;IACH,MAAM,CACJ,KAA6B,EAC7B,OAA+B;QAE/B,MAAM,EAAE,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC;QACzB,MAAM,MAAM,GAAG,IAAI,WAAW,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;QAEhD,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,EAAE,EAAE;YAC3B,OAAO,EAAE,GAAG,EAAE,GAAE,CAAC,EAAE,kCAAkC;YACrD,MAAM,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,GAAG,CAAC;YAClC,MAAM;SACP,CAAC,CAAC;QAEH,IAAI,CAAC,IAAI,CAAC;YACR,IAAI,EAAE,KAAK;YACX,EAAE;YACF,IAAI,EAAE,iBAAiB;YACvB,KAAK;YACL,OAAO,EAAE,OAA8C;SACxD,CAAC,CAAC;QAEH,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,QAAQ,CACZ,KAA6B,EAC7B,OAA+B;QAE/B,MAAM,WAAW,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;QAChD,MAAM,SAAS,GAAG,MAAM,aAAa,CAAC,WAAW,CAAC,CAAC;QACnD,OAAO;YACL,GAAG,SAAS;YACZ,SAAS,EAAE,CAAC,EAAE,oCAAoC;SACnD,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,SAAS;QACP,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,CAAC,CAAC;IACnC,CAAC;IAED;;OAEG;IACH,KAAK;QACH,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC;IAC/B,CAAC;IAED;;;;OAIG;IACH,OAAO,CAAC,QAAiB;QACvB,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,SAAS,EAAE,QAAQ,EAAE,CAAC,CAAC;QACzC,IAAI,QAAQ,EAAE,CAAC;YACb,IAAI,CAAC,YAAY,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QACrC,CAAC;aAAM,CAAC;YACN,IAAI,CAAC,YAAY,CAAC,KAAK,EAAE,CAAC;QAC5B,CAAC;IACH,CAAC;IAED;;OAEG;IACH,QAAQ,CAAC,IAAkB,EAAE,OAAe;QAC1C,OAAO,IAAI,CAAC,YAAY,CAAC,GAAG,CAAC,GAAG,IAAI,KAAK,OAAO,EAAE,CAAC,CAAC;IACtD,CAAC;IAED;;OAEG;IACH,SAAS;QACP,IAAI,CAAC,MAAM,EAAE,SAAS,EAAE,CAAC;QACzB,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC;QACnB,IAAI,CAAC,YAAY,CAAC,KAAK,EAAE,CAAC;QAC1B,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,CAAC;QAC7B,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC;IACzB,CAAC;CACF"}

package/dist/model-worker.d.ts

@@ -1,16 +0,0 @@
-/**
- * Generic Web Worker for running ML pipelines.
- *
- * This file runs in a Web Worker thread. It handles:
- * - WebGPU device detection + WASM fallback
- * - Pipeline instantiation (singleton, keyed by task+model)
- * - Progress reporting during model download
- * - Text generation with streaming + KV cache
- * - One-shot inference for classification, detection, ASR, etc.
- * - GPU device-lost recovery
- * - Memory cleanup
- *
- * Communication is via structured `postMessage` using WorkerCommand/WorkerResponse types.
- */
-export {};
-//# sourceMappingURL=model-worker.d.ts.map

package/dist/model-worker.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"model-worker.d.ts","sourceRoot":"","sources":["../src/model-worker.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG"}

package/dist/model-worker.js

@@ -1,340 +0,0 @@
-/**
- * Generic Web Worker for running ML pipelines.
- *
- * This file runs in a Web Worker thread. It handles:
- * - WebGPU device detection + WASM fallback
- * - Pipeline instantiation (singleton, keyed by task+model)
- * - Progress reporting during model download
- * - Text generation with streaming + KV cache
- * - One-shot inference for classification, detection, ASR, etc.
- * - GPU device-lost recovery
- * - Memory cleanup
- *
- * Communication is via structured `postMessage` using WorkerCommand/WorkerResponse types.
- */
-import { pipeline as hfPipeline, TextStreamer, DynamicCache, InterruptableStoppingCriteria, } from '@huggingface/transformers';
-import { createOnnxPipeline } from './onnx-pipeline.js';
-const instances = new Map();
-const kvCaches = new Map();
-const stoppingCriteria = new InterruptableStoppingCriteria();
-function modelKey(task, modelId) {
-    return `${task}::${modelId}`;
-}
-// ─── Helpers ───
-function send(msg) {
-    self.postMessage(msg);
-}
-// ─── Device Check ───
-async function checkDevice() {
-    let backend = 'cpu';
-    let gpu = null;
-    try {
-        if ('gpu' in navigator) {
-            const adapter = await navigator.gpu.requestAdapter({
-                powerPreference: 'high-performance',
-            });
-            if (adapter) {
-                backend = 'webgpu';
-                const info = adapter.info;
-                const vram = Number(adapter.limits?.maxBufferSize ?? 0);
-                const fmtVram = vram >= 1024 ** 3
-                    ? `${(vram / 1024 ** 3).toFixed(1)} GB`
-                    : `${(vram / 1024 ** 2).toFixed(1)} MB`;
-                gpu = {
-                    vendor: info?.vendor ?? 'unknown',
-                    architecture: info?.architecture ?? 'unknown',
-                    description: info?.description ?? 'unknown',
-                    vram,
-                    vramFormatted: fmtVram,
-                };
-            }
-        }
-    }
-    catch {
-        // WebGPU not available
-    }
-    if (backend === 'cpu' && typeof WebAssembly !== 'undefined') {
-        backend = 'wasm';
-    }
-    // Simple VRAM-based dtype recommendation
-    const vram = gpu?.vram ?? 0;
-    const recommendedDtype = vram >= 8e9 ? 'fp16' : vram >= 4e9 ? 'q8' : 'q4';
-    send({ type: 'device-info', data: { backend, gpu, recommendedDtype } });
-}
-// ─── Load Pipeline ───
-async function loadPipeline(config) {
-    const key = modelKey(config.task, config.modelId);
-    // Dispose previous KV cache if switching models for same task
-    if (kvCaches.has(key)) {
-        kvCaches.get(key)?.dispose?.();
-        kvCaches.delete(key);
-    }
-    send({
-        type: 'progress',
-        data: { status: 'downloading', loaded: 0, total: 0, percent: 0 },
-    });
-    // Create singleton pipeline
-    if (!instances.has(key)) {
-        const isExplicitOnnx = config.modelType === 'onnx' ||
-            config.modelId.endsWith('.onnx') ||
-            config.task === 'raw-onnx' ||
-            config.task === 'custom';
-        const onProgress = (info) => {
-            const loaded = Number(info.loaded ?? 0);
-            const total = Number(info.total ?? 1);
-            const percent = info.percent != null
-                ? Number(info.percent)
-                : (total > 0 ? Math.round((loaded / total) * 100) : 0);
-            send({
-                type: 'progress',
-                data: {
-                    status: info.status ?? 'downloading',
-                    file: String(info.file ?? ''),
-                    loaded,
-                    total,
-                    percent,
-                },
-            });
-        };
-        const loadFn = async () => {
-            if (isExplicitOnnx) {
-                return createOnnxPipeline(config, onProgress);
-            }
-            if (config.modelType === 'transformers') {
-                return (await hfPipeline(config.task, config.modelId, {
-                    device: config.device ?? 'webgpu',
-                    dtype: config.dtype ?? 'q4',
-                    revision: config.revision,
-                    progress_callback: onProgress,
-                }));
-            }
-            // 'auto' mode: try Hugging Face pipeline first, fallback to ONNX if config is missing
-            try {
-                return (await hfPipeline(config.task, config.modelId, {
-                    device: config.device ?? 'webgpu',
-                    dtype: config.dtype ?? 'q4',
-                    revision: config.revision,
-                    progress_callback: onProgress,
-                }));
-            }
-            catch (hfErr) {
-                const errMsg = hfErr instanceof Error ? hfErr.message : String(hfErr);
-                if (errMsg.includes('config.json') ||
-                    errMsg.includes('Could not locate file') ||
-                    errMsg.includes('Unsupported model type')) {
-                    return createOnnxPipeline(config, onProgress);
-                }
-                throw hfErr;
-            }
-        };
-        instances.set(key, loadFn());
-    }
-    try {
-        const generator = await instances.get(key);
-        // Warm up with dummy inference (compiles WebGPU shaders)
-        send({
-            type: 'progress',
-            data: { status: 'compiling', loaded: 0, total: 0, percent: 100 },
-        });
-        if (config.task === 'text-generation' && 'tokenizer' in generator) {
-            const tokenizer = generator.tokenizer;
-            const model = generator.model;
-            if (tokenizer && model && typeof tokenizer === 'function') {
-                const inputs = tokenizer('a');
-                await model.generate({
-                    ...inputs,
-                    max_new_tokens: 1,
-                });
-            }
-        }
-        send({ type: 'ready', modelKey: key });
-    }
-    catch (e) {
-        instances.delete(key);
-        send({
-            type: 'error',
-            id: 'load',
-            data: e instanceof Error ? e.message : String(e),
-        });
-    }
-}
-// ─── Run Inference ───
-async function runInference(id, task, input, options) {
-    // Find the loaded pipeline for this task
-    let pipelineInstance = null;
-    for (const [key, promise] of instances) {
-        if (key.startsWith(`${task}::`)) {
-            pipelineInstance = await promise;
-            break;
-        }
-    }
-    if (!pipelineInstance) {
-        send({ type: 'error', id, data: `No pipeline loaded for task: ${task}` });
-        return;
-    }
-    try {
-        if (task === 'text-generation') {
-            await runTextGeneration(id, pipelineInstance, input, options);
-        }
-        else {
-            // One-shot inference for all other tasks
-            const result = await pipelineInstance(input, options);
-            send({ type: 'result', id, data: result });
-        }
-    }
-    catch (e) {
-        send({
-            type: 'error',
-            id,
-            data: e instanceof Error ? e.message : String(e),
-        });
-    }
-}
-// ─── Text Generation (Streaming) ───
-async function runTextGeneration(id, generator, messages, options) {
-    let startTime = null;
-    let numTokens = 0;
-    let tps = 0;
-    let firstTokenTime = 0;
-    const genObj = generator;
-    if (typeof genObj !== 'function' && !('tokenizer' in genObj)) {
-        send({ type: 'error', id, data: 'Invalid text-generation pipeline' });
-        return;
-    }
-    const streamer = new TextStreamer(genObj.tokenizer, {
-        skip_prompt: true,
-        skip_special_tokens: true,
-        callback_function: (output) => {
-            const event = {
-                token: output,
-                tps,
-                numTokens,
-                timeToFirstToken: firstTokenTime,
-            };
-            send({ type: 'token', id, data: event });
-        },
-        token_callback_function: () => {
-            const now = performance.now();
-            if (startTime === null) {
-                startTime = now;
-                firstTokenTime = now; // Will be relative to request start
-            }
-            numTokens++;
-            if (numTokens > 1 && startTime !== null) {
-                tps = (numTokens / (now - startTime)) * 1000;
-            }
-        },
-    });
-    // Get or create KV cache for this task+model
-    const key = [...instances.keys()].find(k => k.startsWith('text-generation::'));
-    if (key && !kvCaches.has(key)) {
-        kvCaches.set(key, new DynamicCache());
-    }
-    const pastKeyValues = key ? kvCaches.get(key) : undefined;
-    stoppingCriteria.reset();
-    const genOptions = {
-        max_new_tokens: options?.maxNewTokens ?? 1024,
-        do_sample: options?.doSample ?? false,
-        temperature: options?.temperature,
-        top_p: options?.topP,
-        top_k: options?.topK,
-        repetition_penalty: options?.repetitionPenalty,
-        streamer,
-        stopping_criteria: stoppingCriteria,
-        ...(pastKeyValues ? { past_key_values: pastKeyValues } : {}),
-    };
-    // Remove undefined keys
-    for (const k of Object.keys(genOptions)) {
-        if (genOptions[k] === undefined)
-            delete genOptions[k];
-    }
-    try {
-        const output = await genObj(messages, genOptions);
-        // Extract final text
-        let finalText = '';
-        if (Array.isArray(output) && output[0]?.generated_text) {
-            const genText = output[0].generated_text;
-            finalText = Array.isArray(genText)
-                ? genText[genText.length - 1]?.content ?? ''
-                : genText;
-        }
-        send({
-            type: 'result',
-            id,
-            data: {
-                text: finalText,
-                numTokens,
-                tps,
-                timeToFirstToken: firstTokenTime,
-                totalTime: startTime ? performance.now() - startTime : 0,
-            },
-        });
-    }
-    catch (e) {
-        send({
-            type: 'error',
-            id,
-            data: e instanceof Error ? e.message : String(e),
-        });
-    }
-}
-// ─── Cleanup ───
-function disposeModel(targetKey) {
-    if (targetKey) {
-        const p = instances.get(targetKey);
-        p?.then(inst => inst?.dispose?.());
-        instances.delete(targetKey);
-        if (kvCaches.has(targetKey)) {
-            kvCaches.get(targetKey)?.dispose?.();
-            kvCaches.delete(targetKey);
-        }
-    }
-    else {
-        for (const p of instances.values()) {
-            p.then(inst => inst?.dispose?.());
-        }
-        instances.clear();
-        for (const cache of kvCaches.values()) {
-            cache?.dispose?.();
-        }
-        kvCaches.clear();
-    }
-}
-function resetKVCache() {
-    stoppingCriteria.reset();
-    for (const cache of kvCaches.values()) {
-        cache?.dispose?.();
-    }
-    kvCaches.clear();
-}
-// ─── Message Handler ───
-self.addEventListener('message', async (e) => {
-    const cmd = e.data;
-    switch (cmd.type) {
-        case 'check':
-            await checkDevice();
-            break;
-        case 'load':
-            await loadPipeline(cmd.config);
-            break;
-        case 'run':
-            await runInference(cmd.id, cmd.task, cmd.input, cmd.options);
-            break;
-        case 'interrupt':
-            stoppingCriteria.interrupt();
-            break;
-        case 'reset':
-            resetKVCache();
-            break;
-        case 'dispose':
-            disposeModel(cmd.modelKey);
-            break;
-        default:
-            send({
-                type: 'error',
-                id: 'unknown',
-                data: `Unknown command: ${cmd.type}`,
-            });
-    }
-});
-//# sourceMappingURL=model-worker.js.map