@tryhamster/gerbil 1.11.0 → 1.11.2

package/dist/gpu/index.d.mts

@@ -1,3 +1,3 @@
 import { a as OuteSpeakerWord, c as buildOutePromptString, i as OuteSpeaker, l as loadOuteSpeaker, n as OuteSpeakOptions, o as OuteTTS, r as OuteSpeakResult, s as OuteTTSOptions, t as OuteDType } from "../outetts-CAL3_K3j.mjs";
-import { $ as Gemma4VisionGridConfig, A as generateDacSpeechDecoderGraph, At as AdapterSource, B as generateNanoCodecDecoderGraph, Bt as GPUDiagnosticResult, C as TranscribeOptions, Ct as loadKaniTTS, D as generateQwen3_5VisionGraph, Dt as ChatMessage, E as Executor, Et as quantizeKaniBackbone, F as generateMoonshineEncoderGraph, Ft as KaniDType, G as generateGemma4VisionGraph, Gt as ModelArchConfig, H as Gemma4VisionGraphInfo, Ht as GraphDType, I as moonshineEncoderFrames, It as KaniTTS, J as DEFAULT_MODELS, K as patchGemma4VisionClips, Kt as ModelCapabilities, L as parseMoonshineConfig, Lt as KaniTTSOptions, M as parseOuteTtsConfig, Mt as applyLoRAToStore, N as MOONSHINE_REMAINING_WORK, Nt as buildLoRADeltas, O as audioTokensToDacCodes, Ot as Tokenizer, P as generateMoonshineDecoderGraph, Pt as fetchAdapter, Q as GEMMA4_IMAGE_PROCESSOR, R as audioTokensToCodes, Rt as SpeakOptions, S as MoonshineSTTOptions, St as LoadedMoonshine, T as MoonshineEncoderExecutor, Tt as loadMoonshine, U as dequantizeGemma4VisionProjection, Ut as KVDType, V as parseKaniConfig, Vt as initGPU, W as dequantizeMLXProjection, Wt as KvMode, X as OUTETTS_PRESET_VOICES, Y as OUTETTS_ASSETS, Z as resolveDefaultRepo, _ as ParlerSpeakOptions, _t as preprocessImage, a as GenerateObjectOptions, at as VisionPositionTensors, b as ParlerTTSOptions, bt as SamplingParams, c as GenerateResult, ct as buildGemma4RotaryCosSin, d as ObjectSchema, dt as buildMRoPEPositionIds, et as Gemma4VisionPositionTensors, f as ObjectValidator, ft as buildPosEmbeds, g as VisionInputs, gt as mropeFreqDims, h as VisionExecutor, ht as buildVisionPositionTensors, i as EncodeImageResult, it as VisionGridConfig, j as generateOuteTtsBackboneGraph, jt as LoRADelta, k as dacOutputLength, kt as AdapterConfig, l as IntegrityCheckEntry, lt as buildGemma4VisionPositionTensors, m as WebGPUEngineOptions, mt as buildRotaryCosSin, n as AgentTool, nt as PreprocessedImage, o as GenerateObjectResult, ot as buildGemma4PoolMatrix, p as WebGPUEngine, pt as buildPositionIds, q as resolveGemma4VisionInfo, qt as createDefaultHFKeyMapper, r as EmbedOptions, rt as QWEN3_5_IMAGE_PROCESSOR, s as GenerateOptions, st as buildGemma4PosEmbeds, t as AgentStep, tt as ImageProcessorConfig, u as IntegrityCheckResult, ut as buildMRoPECosSin, v as ParlerSpeakResult, vt as preprocessImageGemma4, w as TranscribeResult, wt as loadModel, x as MoonshineSTT, xt as LoadedKaniTTS, y as ParlerTTS, yt as smartResize, z as generateKaniTtsGraph, zt as SpeakResult } from "../index-DyqcKFfQ.mjs";
-export { AdapterConfig, AdapterSource, AgentStep, AgentTool, ChatMessage, DEFAULT_MODELS, EmbedOptions, EncodeImageResult, Executor, GEMMA4_IMAGE_PROCESSOR, GPUDiagnosticResult, Gemma4VisionGraphInfo, Gemma4VisionGridConfig, Gemma4VisionPositionTensors, GenerateObjectOptions, GenerateObjectResult, GenerateOptions, GenerateResult, GraphDType, ImageProcessorConfig, IntegrityCheckEntry, IntegrityCheckResult, KVDType, KaniDType, KaniTTS, KaniTTSOptions, KvMode, LoRADelta, LoadedKaniTTS, LoadedMoonshine, MOONSHINE_REMAINING_WORK, ModelArchConfig, ModelCapabilities, MoonshineEncoderExecutor, MoonshineSTT, MoonshineSTTOptions, OUTETTS_ASSETS, OUTETTS_PRESET_VOICES, ObjectSchema, ObjectValidator, OuteDType, OuteSpeakOptions, OuteSpeakResult, OuteSpeaker, OuteSpeakerWord, OuteTTS, OuteTTSOptions, ParlerSpeakOptions, ParlerSpeakResult, ParlerTTS, ParlerTTSOptions, PreprocessedImage, QWEN3_5_IMAGE_PROCESSOR, SamplingParams, SpeakOptions, SpeakResult, Tokenizer, TranscribeOptions, TranscribeResult, VisionExecutor, VisionGridConfig, VisionInputs, VisionPositionTensors, WebGPUEngine, WebGPUEngineOptions, applyLoRAToStore, audioTokensToCodes, audioTokensToDacCodes, buildGemma4PoolMatrix, buildGemma4PosEmbeds, buildGemma4RotaryCosSin, buildGemma4VisionPositionTensors, buildLoRADeltas, buildMRoPECosSin, buildMRoPEPositionIds, buildOutePromptString, buildPosEmbeds, buildPositionIds, buildRotaryCosSin, buildVisionPositionTensors, createDefaultHFKeyMapper, dacOutputLength, dequantizeGemma4VisionProjection, dequantizeMLXProjection, fetchAdapter, generateDacSpeechDecoderGraph, generateGemma4VisionGraph, generateKaniTtsGraph, generateMoonshineDecoderGraph, generateMoonshineEncoderGraph, generateNanoCodecDecoderGraph, generateOuteTtsBackboneGraph, generateQwen3_5VisionGraph, initGPU, loadKaniTTS, loadModel, loadMoonshine, loadOuteSpeaker, moonshineEncoderFrames, mropeFreqDims, parseKaniConfig, parseMoonshineConfig, parseOuteTtsConfig, patchGemma4VisionClips, preprocessImage, preprocessImageGemma4, quantizeKaniBackbone, resolveDefaultRepo, resolveGemma4VisionInfo, smartResize };
+import { $ as GEMMA4_IMAGE_PROCESSOR, A as generateDacSpeechDecoderGraph, At as AdapterConfig, B as generateNanoCodecDecoderGraph, Bt as SpeakOptions, C as TranscribeOptions, Ct as LoadedMoonshine, D as generateQwen3_5VisionGraph, Dt as quantizeKaniBackbone, E as Executor, Et as loadMoonshine, F as generateMoonshineEncoderGraph, Ft as fetchAdapter, G as generateGemma4VisionGraph, Gt as KVDType, H as Gemma4VisionGraphInfo, Ht as GPUDiagnosticResult, I as moonshineEncoderFrames, It as splitFusedQGateDelta, J as DEFAULT_MODELS, Jt as ModelCapabilities, K as patchGemma4VisionClips, Kt as KvMode, L as parseMoonshineConfig, Lt as KaniDType, M as parseOuteTtsConfig, Mt as LoRADelta, N as MOONSHINE_REMAINING_WORK, Nt as applyLoRAToStore, O as audioTokensToDacCodes, Ot as ChatMessage, P as generateMoonshineDecoderGraph, Pt as buildLoRADeltas, Q as resolveDefaultRepo, R as audioTokensToCodes, Rt as KaniTTS, S as MoonshineSTTOptions, St as LoadedKaniTTS, T as MoonshineEncoderExecutor, Tt as loadModel, U as dequantizeGemma4VisionProjection, Ut as initGPU, V as parseKaniConfig, Vt as SpeakResult, W as dequantizeMLXProjection, Wt as GraphDType, X as OUTETTS_PRESET_VOICES, Y as OUTETTS_ASSETS, Yt as createDefaultHFKeyMapper, Z as isTTSRepo, _ as ParlerSpeakOptions, _t as mropeFreqDims, a as GenerateObjectOptions, at as VisionGridConfig, b as ParlerTTSOptions, bt as smartResize, c as GenerateResult, ct as buildGemma4PosEmbeds, d as ObjectSchema, dt as buildMRoPECosSin, et as Gemma4VisionGridConfig, f as ObjectValidator, ft as buildMRoPEPositionIds, g as VisionInputs, gt as buildVisionPositionTensors, h as VisionExecutor, ht as buildRotaryCosSin, i as EncodeImageResult, it as QWEN3_5_IMAGE_PROCESSOR, j as generateOuteTtsBackboneGraph, jt as AdapterSource, k as dacOutputLength, kt as Tokenizer, l as IntegrityCheckEntry, lt as buildGemma4RotaryCosSin, m as WebGPUEngineOptions, mt as buildPositionIds, n as AgentTool, nt as ImageProcessorConfig, o as GenerateObjectResult, ot as VisionPositionTensors, p as WebGPUEngine, pt as buildPosEmbeds, q as resolveGemma4VisionInfo, qt as ModelArchConfig, r as EmbedOptions, rt as PreprocessedImage, s as GenerateOptions, st as buildGemma4PoolMatrix, t as AgentStep, tt as Gemma4VisionPositionTensors, u as IntegrityCheckResult, ut as buildGemma4VisionPositionTensors, v as ParlerSpeakResult, vt as preprocessImage, w as TranscribeResult, wt as loadKaniTTS, x as MoonshineSTT, xt as SamplingParams, y as ParlerTTS, yt as preprocessImageGemma4, z as generateKaniTtsGraph, zt as KaniTTSOptions } from "../index-t_hY-XWW.mjs";
+export { AdapterConfig, AdapterSource, AgentStep, AgentTool, ChatMessage, DEFAULT_MODELS, EmbedOptions, EncodeImageResult, Executor, GEMMA4_IMAGE_PROCESSOR, GPUDiagnosticResult, Gemma4VisionGraphInfo, Gemma4VisionGridConfig, Gemma4VisionPositionTensors, GenerateObjectOptions, GenerateObjectResult, GenerateOptions, GenerateResult, GraphDType, ImageProcessorConfig, IntegrityCheckEntry, IntegrityCheckResult, KVDType, KaniDType, KaniTTS, KaniTTSOptions, KvMode, LoRADelta, LoadedKaniTTS, LoadedMoonshine, MOONSHINE_REMAINING_WORK, ModelArchConfig, ModelCapabilities, MoonshineEncoderExecutor, MoonshineSTT, MoonshineSTTOptions, OUTETTS_ASSETS, OUTETTS_PRESET_VOICES, ObjectSchema, ObjectValidator, OuteDType, OuteSpeakOptions, OuteSpeakResult, OuteSpeaker, OuteSpeakerWord, OuteTTS, OuteTTSOptions, ParlerSpeakOptions, ParlerSpeakResult, ParlerTTS, ParlerTTSOptions, PreprocessedImage, QWEN3_5_IMAGE_PROCESSOR, SamplingParams, SpeakOptions, SpeakResult, Tokenizer, TranscribeOptions, TranscribeResult, VisionExecutor, VisionGridConfig, VisionInputs, VisionPositionTensors, WebGPUEngine, WebGPUEngineOptions, applyLoRAToStore, audioTokensToCodes, audioTokensToDacCodes, buildGemma4PoolMatrix, buildGemma4PosEmbeds, buildGemma4RotaryCosSin, buildGemma4VisionPositionTensors, buildLoRADeltas, buildMRoPECosSin, buildMRoPEPositionIds, buildOutePromptString, buildPosEmbeds, buildPositionIds, buildRotaryCosSin, buildVisionPositionTensors, createDefaultHFKeyMapper, dacOutputLength, dequantizeGemma4VisionProjection, dequantizeMLXProjection, fetchAdapter, generateDacSpeechDecoderGraph, generateGemma4VisionGraph, generateKaniTtsGraph, generateMoonshineDecoderGraph, generateMoonshineEncoderGraph, generateNanoCodecDecoderGraph, generateOuteTtsBackboneGraph, generateQwen3_5VisionGraph, initGPU, isTTSRepo, loadKaniTTS, loadModel, loadMoonshine, loadOuteSpeaker, moonshineEncoderFrames, mropeFreqDims, parseKaniConfig, parseMoonshineConfig, parseOuteTtsConfig, patchGemma4VisionClips, preprocessImage, preprocessImageGemma4, quantizeKaniBackbone, resolveDefaultRepo, resolveGemma4VisionInfo, smartResize, splitFusedQGateDelta };

package/dist/gpu/index.mjs

@@ -1,6 +1,6 @@
-import { i as resolveDefaultRepo, n as OUTETTS_ASSETS, r as OUTETTS_PRESET_VOICES, t as DEFAULT_MODELS } from "../defaults-C_bJK9zs.mjs";
+import { a as resolveDefaultRepo, i as isTTSRepo, n as OUTETTS_ASSETS, r as OUTETTS_PRESET_VOICES, t as DEFAULT_MODELS } from "../defaults-DfGx4d1m.mjs";
 import { E as generateNanoCodecDecoderGraph, M as parseKaniConfig, R as createDefaultHFKeyMapper, T as generateKaniTtsGraph, _ as moonshineEncoderFrames, g as generateMoonshineEncoderGraph, h as generateMoonshineDecoderGraph, m as MOONSHINE_REMAINING_WORK, v as parseMoonshineConfig, x as audioTokensToCodes } from "../architectures-DmZMEFsA.mjs";
-import { A as dequantizeGemma4VisionProjection, C as audioTokensToDacCodes, D as parseOuteTtsConfig, E as generateOuteTtsBackboneGraph, M as generateGemma4VisionGraph, N as patchGemma4VisionClips, O as KaniTTS, P as resolveGemma4VisionInfo, S as loadOuteSpeaker, T as generateDacSpeechDecoderGraph, _ as smartResize, a as buildGemma4PosEmbeds, b as OuteTTS, c as buildMRoPECosSin, d as buildPositionIds, f as buildRotaryCosSin, g as preprocessImageGemma4, h as preprocessImage, i as buildGemma4PoolMatrix, j as dequantizeMLXProjection, k as generateQwen3_5VisionGraph, l as buildMRoPEPositionIds, m as mropeFreqDims, n as GEMMA4_IMAGE_PROCESSOR, o as buildGemma4RotaryCosSin, p as buildVisionPositionTensors, r as QWEN3_5_IMAGE_PROCESSOR, s as buildGemma4VisionPositionTensors, t as WebGPUEngine, u as buildPosEmbeds, v as VisionExecutor, w as dacOutputLength, x as buildOutePromptString, y as ParlerTTS } from "../gpu-CzgbyVeq.mjs";
-import { T as initGPU, a as loadModel, f as Tokenizer, g as Executor, h as fetchAdapter, i as loadKaniTTS, m as buildLoRADeltas, n as MoonshineEncoderExecutor, o as loadMoonshine, p as applyLoRAToStore, t as MoonshineSTT, u as quantizeKaniBackbone } from "../moonshine-stt-CgDXoLFd.mjs";
+import { A as dequantizeGemma4VisionProjection, C as audioTokensToDacCodes, D as parseOuteTtsConfig, E as generateOuteTtsBackboneGraph, M as generateGemma4VisionGraph, N as patchGemma4VisionClips, O as KaniTTS, P as resolveGemma4VisionInfo, S as loadOuteSpeaker, T as generateDacSpeechDecoderGraph, _ as smartResize, a as buildGemma4PosEmbeds, b as OuteTTS, c as buildMRoPECosSin, d as buildPositionIds, f as buildRotaryCosSin, g as preprocessImageGemma4, h as preprocessImage, i as buildGemma4PoolMatrix, j as dequantizeMLXProjection, k as generateQwen3_5VisionGraph, l as buildMRoPEPositionIds, m as mropeFreqDims, n as GEMMA4_IMAGE_PROCESSOR, o as buildGemma4RotaryCosSin, p as buildVisionPositionTensors, r as QWEN3_5_IMAGE_PROCESSOR, s as buildGemma4VisionPositionTensors, t as WebGPUEngine, u as buildPosEmbeds, v as VisionExecutor, w as dacOutputLength, x as buildOutePromptString, y as ParlerTTS } from "../gpu-BQ8I-VDx.mjs";
+import { E as initGPU, _ as splitFusedQGateDelta, a as loadModel, f as Tokenizer, g as fetchAdapter, h as buildLoRADeltas, i as loadKaniTTS, m as applyLoRAToStore, n as MoonshineEncoderExecutor, o as loadMoonshine, p as Executor, t as MoonshineSTT, u as quantizeKaniBackbone } from "../moonshine-stt-BI-5UmD_.mjs";
 
-export { DEFAULT_MODELS, Executor, GEMMA4_IMAGE_PROCESSOR, KaniTTS, MOONSHINE_REMAINING_WORK, MoonshineEncoderExecutor, MoonshineSTT, OUTETTS_ASSETS, OUTETTS_PRESET_VOICES, OuteTTS, ParlerTTS, QWEN3_5_IMAGE_PROCESSOR, Tokenizer, VisionExecutor, WebGPUEngine, applyLoRAToStore, audioTokensToCodes, audioTokensToDacCodes, buildGemma4PoolMatrix, buildGemma4PosEmbeds, buildGemma4RotaryCosSin, buildGemma4VisionPositionTensors, buildLoRADeltas, buildMRoPECosSin, buildMRoPEPositionIds, buildOutePromptString, buildPosEmbeds, buildPositionIds, buildRotaryCosSin, buildVisionPositionTensors, createDefaultHFKeyMapper, dacOutputLength, dequantizeGemma4VisionProjection, dequantizeMLXProjection, fetchAdapter, generateDacSpeechDecoderGraph, generateGemma4VisionGraph, generateKaniTtsGraph, generateMoonshineDecoderGraph, generateMoonshineEncoderGraph, generateNanoCodecDecoderGraph, generateOuteTtsBackboneGraph, generateQwen3_5VisionGraph, initGPU, loadKaniTTS, loadModel, loadMoonshine, loadOuteSpeaker, moonshineEncoderFrames, mropeFreqDims, parseKaniConfig, parseMoonshineConfig, parseOuteTtsConfig, patchGemma4VisionClips, preprocessImage, preprocessImageGemma4, quantizeKaniBackbone, resolveDefaultRepo, resolveGemma4VisionInfo, smartResize };
+export { DEFAULT_MODELS, Executor, GEMMA4_IMAGE_PROCESSOR, KaniTTS, MOONSHINE_REMAINING_WORK, MoonshineEncoderExecutor, MoonshineSTT, OUTETTS_ASSETS, OUTETTS_PRESET_VOICES, OuteTTS, ParlerTTS, QWEN3_5_IMAGE_PROCESSOR, Tokenizer, VisionExecutor, WebGPUEngine, applyLoRAToStore, audioTokensToCodes, audioTokensToDacCodes, buildGemma4PoolMatrix, buildGemma4PosEmbeds, buildGemma4RotaryCosSin, buildGemma4VisionPositionTensors, buildLoRADeltas, buildMRoPECosSin, buildMRoPEPositionIds, buildOutePromptString, buildPosEmbeds, buildPositionIds, buildRotaryCosSin, buildVisionPositionTensors, createDefaultHFKeyMapper, dacOutputLength, dequantizeGemma4VisionProjection, dequantizeMLXProjection, fetchAdapter, generateDacSpeechDecoderGraph, generateGemma4VisionGraph, generateKaniTtsGraph, generateMoonshineDecoderGraph, generateMoonshineEncoderGraph, generateNanoCodecDecoderGraph, generateOuteTtsBackboneGraph, generateQwen3_5VisionGraph, initGPU, isTTSRepo, loadKaniTTS, loadModel, loadMoonshine, loadOuteSpeaker, moonshineEncoderFrames, mropeFreqDims, parseKaniConfig, parseMoonshineConfig, parseOuteTtsConfig, patchGemma4VisionClips, preprocessImage, preprocessImageGemma4, quantizeKaniBackbone, resolveDefaultRepo, resolveGemma4VisionInfo, smartResize, splitFusedQGateDelta };

package/dist/{gpu-CzgbyVeq.mjs → gpu-BQ8I-VDx.mjs}

@@ -1,6 +1,6 @@
-import { i as resolveDefaultRepo, n as OUTETTS_ASSETS, r as OUTETTS_PRESET_VOICES, t as DEFAULT_MODELS } from "./defaults-C_bJK9zs.mjs";
+import { a as resolveDefaultRepo, i as isTTSRepo, n as OUTETTS_ASSETS, r as OUTETTS_PRESET_VOICES, t as DEFAULT_MODELS } from "./defaults-DfGx4d1m.mjs";
 import { A as kaniSinTensor, C as computeKaniPositions, D as kaniAttentionLayerIndices, E as generateNanoCodecDecoderGraph, F as CANONICAL_KEYS, I as DTYPE_BYTES, L as GEMMA4_VIS_KEYS, M as parseKaniConfig, N as DEFAULT_GROUP_SIZE, O as kaniCosTensor, S as buildKaniLayerCosSin, T as generateKaniTtsGraph, a as PARLER_DAC_LATENT_DIM, b as KANI_START_OF_HUMAN, c as PARLER_SAMPLE_RATE, d as generateParlerEncoderGraph, f as parseParlerConfig, i as PARLER_DAC_DECODER_DIM, k as kaniLayerAlpha, l as buildT5RelativeBias, o as PARLER_DECODER_RATES, p as revertDelayPattern, r as PARLER_BOS_TOKEN_ID, s as PARLER_EOS_TOKEN_ID, u as generateParlerDecoderGraph, x as audioTokensToCodes, y as KANI_END_OF_HUMAN } from "./architectures-DmZMEFsA.mjs";
-import { C as destroyBuffers, E as verifyGPU, S as createUniformBuffer, T as initGPU, _ as KERNEL_REGISTRY, a as loadModel, b as createBindGroup, c as loadParlerTTS, d as remapPrunedToken, g as Executor, h as fetchAdapter, i as loadKaniTTS, l as quantizeBackboneInt4, m as buildLoRADeltas, r as createKeyMapperForArch, s as loadOuteTTS, u as quantizeKaniBackbone, v as MATMUL_BIAS_F16C_SPEC, w as getOrCreatePipeline, x as createStorageBuffer, y as clearPipelineCache } from "./moonshine-stt-CgDXoLFd.mjs";
+import { C as createUniformBuffer, D as verifyGPU, E as initGPU, S as createStorageBuffer, T as getOrCreatePipeline, a as loadModel, b as clearPipelineCache, c as loadParlerTTS, d as remapPrunedToken, g as fetchAdapter, h as buildLoRADeltas, i as loadKaniTTS, l as quantizeBackboneInt4, p as Executor, r as createKeyMapperForArch, s as loadOuteTTS, u as quantizeKaniBackbone, v as KERNEL_REGISTRY, w as destroyBuffers, x as createBindGroup, y as MATMUL_BIAS_F16C_SPEC } from "./moonshine-stt-BI-5UmD_.mjs";
 
 //#region src/gpu/architectures/gemma4_vision.ts
 /**
@@ -5101,9 +5101,22 @@ function parseAgentToolCall(text) {
 	return null;
 }
 var WebGPUEngine = class WebGPUEngine {
-	ctx;
-	executor;
-	tokenizer;
+	_ctx;
+	_executor;
+	_tokenizer;
+	get ctx() {
+		if (!this._ctx) throw new Error(WebGPUEngine.TTS_LITE_ERROR);
+		return this._ctx;
+	}
+	get executor() {
+		if (!this._executor) throw new Error(WebGPUEngine.TTS_LITE_ERROR);
+		return this._executor;
+	}
+	get tokenizer() {
+		if (!this._tokenizer) throw new Error(WebGPUEngine.TTS_LITE_ERROR);
+		return this._tokenizer;
+	}
+	static TTS_LITE_ERROR = "This engine was created for a TTS checkpoint; base-model ops (generate/embed/describeImage) are unavailable. Use speak() / encodeSpeaker().";
 	_destroyed = false;
 	_isEmbedding;
 	/** HF architecture string (e.g. "Gemma3TextModel", "Qwen3ForCausalLM"). */
@@ -5155,16 +5168,25 @@ var WebGPUEngine = class WebGPUEngine {
 	_genQueueTail = Promise.resolve();
 	/** Model capabilities (text, vision, moe). */
 	capabilities;
-	/** Model architecture config. */
-	config;
+	/** Model architecture config (null on a TTS-only lite engine). */
+	_config;
+	/** Model architecture config. Throws on a TTS-only lite engine. */
+	get config() {
+		if (!this._config) throw new Error(WebGPUEngine.TTS_LITE_ERROR);
+		return this._config;
+	}
 	constructor(ctx, executor, tokenizer, graph, opts, vision) {
-		this.ctx = ctx;
-		this.executor = executor;
-		this.tokenizer = tokenizer;
-		this.capabilities = graph.capabilities;
-		this.config = graph.config;
-		this._isEmbedding = graph.outputs.includes("embedding");
-		this._architecture = graph.architecture;
+		this._ctx = ctx;
+		this._executor = executor;
+		this._tokenizer = tokenizer;
+		this.capabilities = graph?.capabilities ?? {
+			text: true,
+			vision: false,
+			moe: false
+		};
+		this._config = graph?.config ?? null;
+		this._isEmbedding = graph?.outputs.includes("embedding") ?? false;
+		this._architecture = graph?.architecture ?? "";
 		this.visionExecutor = vision?.executor ?? null;
 		this.visionConfig = vision?.config ?? null;
 		this.visionPosEmbedTable = vision?.posEmbedTable ?? null;
@@ -5281,6 +5303,7 @@ var WebGPUEngine = class WebGPUEngine {
 			...options,
 			repo: resolveDefaultRepo(options)
 		};
+		if (isTTSRepo(options.repo)) return WebGPUEngine._buildTTSLite(options);
 		const ctx = await initGPU();
 		try {
 			return await WebGPUEngine._buildFromContext(ctx, options);
@@ -5291,6 +5314,21 @@ var WebGPUEngine = class WebGPUEngine {
 			throw err;
 		}
 	}
+	/**
+	* Build a lightweight TTS-only engine: no GPU device is initialized and no base
+	* text graph is loaded. It stores the create() options; the first `speak()`
+	* (or `encodeSpeaker()`) lazily builds the dedicated native TTS engine —
+	* {@link KaniTTS}, {@link OuteTTS}, or {@link ParlerTTS} — which brings its own
+	* GPU device and weights. Base-model ops throw {@link WebGPUEngine.TTS_LITE_ERROR}.
+	*/
+	static _buildTTSLite(options) {
+		return new WebGPUEngine(null, null, null, null, {
+			multimodalGraph: false,
+			rawConfig: {},
+			maxSeqLen: options.maxSeqLen ?? 2048,
+			createOptions: options
+		}, null);
+	}
 	static async _buildFromContext(ctx, options) {
 		const isBrowser = typeof navigator !== "undefined" && typeof location !== "undefined";
 		const isSafari = isBrowser && /Safari/.test(navigator.userAgent) && !/Chrome/.test(navigator.userAgent);
@@ -6685,13 +6723,15 @@ var WebGPUEngine = class WebGPUEngine {
 	destroy() {
 		if (this._destroyed) return;
 		this._destroyed = true;
-		this.executor.destroy();
+		this._executor?.destroy();
 		this._kaniTTS?.destroy();
 		this._outeTTS?.destroy();
 		this._parlerTTS?.destroy();
 		this.visionExecutor?.destroy();
-		clearPipelineCache(this.ctx.device);
-		this.ctx.device.destroy();
+		if (this._ctx) {
+			clearPipelineCache(this._ctx.device);
+			this._ctx.device.destroy();
+		}
 	}
 	checkDestroyed() {
 		if (this._destroyed) throw new Error("WebGPUEngine has been destroyed");
@@ -6722,4 +6762,4 @@ var WebGPUEngine = class WebGPUEngine {
 
 //#endregion
 export { dequantizeGemma4VisionProjection as A, audioTokensToDacCodes as C, parseOuteTtsConfig as D, generateOuteTtsBackboneGraph as E, generateGemma4VisionGraph as M, patchGemma4VisionClips as N, KaniTTS as O, resolveGemma4VisionInfo as P, loadOuteSpeaker as S, generateDacSpeechDecoderGraph as T, smartResize as _, buildGemma4PosEmbeds as a, OuteTTS as b, buildMRoPECosSin as c, buildPositionIds as d, buildRotaryCosSin as f, preprocessImageGemma4 as g, preprocessImage as h, buildGemma4PoolMatrix as i, dequantizeMLXProjection as j, generateQwen3_5VisionGraph as k, buildMRoPEPositionIds as l, mropeFreqDims as m, GEMMA4_IMAGE_PROCESSOR as n, buildGemma4RotaryCosSin as o, buildVisionPositionTensors as p, QWEN3_5_IMAGE_PROCESSOR as r, buildGemma4VisionPositionTensors as s, WebGPUEngine as t, buildPosEmbeds as u, VisionExecutor as v, dacOutputLength as w, buildOutePromptString as x, ParlerTTS as y };
-//# sourceMappingURL=gpu-CzgbyVeq.mjs.map
+//# sourceMappingURL=gpu-BQ8I-VDx.mjs.map