@tryhamster/gerbil 1.8.0 → 1.10.0

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-ElJKy9i9.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-B56xBiR2.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 };

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 { D as createDefaultHFKeyMapper, a as generateMoonshineEncoderGraph, h as generateNanoCodecDecoderGraph, i as generateMoonshineDecoderGraph, m as generateKaniTtsGraph, o as moonshineEncoderFrames, r as MOONSHINE_REMAINING_WORK, s as parseMoonshineConfig, u as audioTokensToCodes, x as parseKaniConfig } from "../architectures-BHkqQ9xp.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-kQVLpV3n.mjs";
-import { a as loadMoonshine, d as Tokenizer, f as applyLoRAToStore, h as Executor, i as loadModel, l as quantizeKaniBackbone, m as fetchAdapter, n as MoonshineEncoderExecutor, p as buildLoRADeltas, r as loadKaniTTS, t as MoonshineSTT, w as initGPU } from "../moonshine-stt-9wc6t11v.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-NPdk7pmr.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-BYbYwSz2.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 };

package/dist/{gpu-kQVLpV3n.mjs → gpu-NPdk7pmr.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 { E as GEMMA4_VIS_KEYS, S as DEFAULT_GROUP_SIZE, T as DTYPE_BYTES, _ as kaniCosTensor, c as KANI_END_OF_HUMAN, d as buildKaniLayerCosSin, f as computeKaniPositions, g as kaniAttentionLayerIndices, h as generateNanoCodecDecoderGraph, l as KANI_START_OF_HUMAN, m as generateKaniTtsGraph, u as audioTokensToCodes, v as kaniLayerAlpha, w as CANONICAL_KEYS, x as parseKaniConfig, y as kaniSinTensor } from "./architectures-BHkqQ9xp.mjs";
-import { C as getOrCreatePipeline, S as destroyBuffers, T as verifyGPU, _ as MATMUL_BIAS_F16C_SPEC, b as createStorageBuffer, c as quantizeBackboneInt4, g as KERNEL_REGISTRY, h as Executor, i as loadModel, l as quantizeKaniBackbone, o as loadOuteTTS, r as loadKaniTTS, s as loadParlerTTS, u as remapPrunedToken, v as clearPipelineCache, w as initGPU, x as createUniformBuffer, y as createBindGroup } from "./moonshine-stt-9wc6t11v.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-BYbYwSz2.mjs";
 
 //#region src/gpu/architectures/gemma4_vision.ts
 /**
@@ -1123,6 +1123,108 @@ var KaniTTS = class KaniTTS {
 	}
 };
 
+//#endregion
+//#region src/gpu/autocomplete-clean.ts
+const DEFAULT_MAX_SUGGESTION_CHARS = 140;
+const DEFAULT_RESTATEMENT_MIN_WORDS = 3;
+const DEFAULT_INTERNAL_REPEAT_MIN_RUN = 6;
+const LEADING_PUNCT = /^[.,;:!?)\]}'"”’%]/;
+const WRAPPING_QUOTES_START = /^["'“”‘’`]+/;
+const WRAPPING_QUOTES_END = /["'“”‘’`]+$/;
+const TRAILING_WHITESPACE = /\s+$/;
+const WHITESPACE_RUN = /\s+/g;
+const AFTER_FIRST_NEWLINE = /\r?\n[\s\S]*$/;
+const ENDS_WITH_SPACE = /\s$/;
+/**
+* Longest span of characters that is both a suffix of `typed` and a prefix of
+* `suggestion`, dropped from the start of `suggestion`. Trailing whitespace on
+* `typed` is ignored so a duplicated word at the cursor ("the " + "the store")
+* is caught. Returns `suggestion` unchanged when there is no overlap.
+*/
+function stripLeadingOverlap(typed, suggestion) {
+	const t = typed.replace(TRAILING_WHITESPACE, "");
+	const max = Math.min(t.length, suggestion.length);
+	for (let k = max; k > 0; k--) if (t.slice(t.length - k) === suggestion.slice(0, k)) return suggestion.slice(k);
+	return suggestion;
+}
+/**
+* Drop a leading run of words from `suggestion` when that run already appears as
+* a contiguous phrase anywhere in `typed`. This catches the model restating an
+* EARLIER phrase (not just the immediate tail). Matching is case-insensitive,
+* whitespace-normalized, and word-boundary aware so it never strips on a
+* partial-word coincidence. Only strips runs of at least `minWords` words to
+* avoid nuking a legitimate short continuation.
+*/
+function stripRestatement(typed, suggestion, minWords = DEFAULT_RESTATEMENT_MIN_WORDS) {
+	const normTyped = ` ${typed.toLowerCase().replace(WHITESPACE_RUN, " ").trim()} `;
+	const words = suggestion.split(WHITESPACE_RUN).filter(Boolean);
+	let matched = 0;
+	for (let n = 1; n <= words.length; n++) {
+		const phrase = ` ${words.slice(0, n).join(" ").toLowerCase()} `;
+		if (normTyped.includes(phrase)) matched = n;
+		else break;
+	}
+	if (matched >= minWords) return words.slice(matched).join(" ");
+	return suggestion;
+}
+/**
+* Truncate `text` right before the first point where a span of `minRun` or more
+* consecutive words repeats a span seen earlier — the classic small-model loop
+* ("…endless sands …endless sands…"). Returns `text` unchanged (spacing
+* preserved) when no such repeat exists.
+*/
+function truncateInternalRepeat(text, minRun = DEFAULT_INTERNAL_REPEAT_MIN_RUN) {
+	const words = text.split(WHITESPACE_RUN).filter(Boolean);
+	const seen = /* @__PURE__ */ new Set();
+	for (let j = 0; j + minRun <= words.length; j++) {
+		const gram = words.slice(j, j + minRun).join(" ").toLowerCase();
+		if (seen.has(gram)) return words.slice(0, j).join(" ");
+		seen.add(gram);
+	}
+	return text;
+}
+/**
+* Cap `text` to at most `max` characters, cutting on a word boundary when a
+* reasonable one exists in the back half of the window.
+*/
+function capLength(text, max) {
+	if (text.length <= max) return text;
+	const cut = text.slice(0, max);
+	const lastSpace = cut.lastIndexOf(" ");
+	return (lastSpace > max * .5 ? cut.slice(0, lastSpace) : cut).trimEnd();
+}
+/**
+* Turn a raw model completion into a clean inline ghost continuation:
+* 1. keep only the first line (when `singleLine`),
+* 2. strip wrapping quotes,
+* 3. drop a full verbatim echo of the typed text,
+* 4. drop a leading character overlap with the typed tail,
+* 5. drop a leading word-run that restates an earlier typed phrase,
+* 6. truncate an internal phrase loop,
+* 7. cap the length,
+* 8. add a single smart leading space so the ghost joins the caret naturally.
+*
+* Returns "" when nothing novel is left — the ghost then simply shows nothing,
+* which is the correct behavior for a suggestion that only repeats the input.
+*/
+function cleanSuggestion(raw, typed, options = {}) {
+	const { singleLine = true, maxChars = DEFAULT_MAX_SUGGESTION_CHARS } = options;
+	let s = singleLine ? raw.replace(AFTER_FIRST_NEWLINE, "") : raw;
+	s = s.replace(WRAPPING_QUOTES_START, "").replace(WRAPPING_QUOTES_END, "");
+	s = s.trim();
+	if (!s) return "";
+	const typedTrim = typed.trim();
+	if (typedTrim && s.startsWith(typedTrim)) s = s.slice(typedTrim.length).trimStart();
+	s = stripLeadingOverlap(typed, s).trimStart();
+	s = stripRestatement(typed, s);
+	s = truncateInternalRepeat(s);
+	s = capLength(s.trim(), maxChars).trim();
+	if (!s) return "";
+	const startsWithPunct = LEADING_PUNCT.test(s);
+	const typedEndsWithSpace = ENDS_WITH_SPACE.test(typed) || typed.length === 0;
+	return startsWithPunct || typedEndsWithSpace ? s : ` ${s}`;
+}
+
 //#endregion
 //#region src/gpu/architectures/outetts.ts
 const OUTETTS_C1_BASE = 151669;
@@ -5516,22 +5618,6 @@ const AUTOCOMPLETE_SYSTEM = [
 	"Do not answer questions; just continue the writing.",
 	"Example — input: \"The quick brown fox\" → continuation: \" jumps over the lazy dog.\""
 ].join(" ");
-/**
-* Turn raw model output into a clean inline continuation: cut after the first
-* newline (single-line), strip wrapping quotes, drop an echoed copy of the typed
-* text, and add a single leading space unless the suggestion hugs punctuation or
-* the typed text already ends with whitespace.
-*/
-function normalizeContinuation(raw, typed, singleLine) {
-	let s = singleLine ? raw.replace(/\n[\s\S]*$/, "") : raw;
-	s = s.replace(/^["'“”']+/, "").replace(/["'“”']+$/, "");
-	if (s.startsWith(typed)) s = s.slice(typed.length);
-	s = s.replace(/^\s+/, "");
-	if (!s) return "";
-	const startsWithPunct = /^[.,;:!?)\]}'"”’%]/.test(s);
-	const typedEndsWithSpace = /\s$/.test(typed) || typed.length === 0;
-	return startsWithPunct || typedEndsWithSpace ? s : ` ${s}`;
-}
 function formatAgentToolsPrompt(tools) {
 	return `You are a helpful assistant with access to tools.
 
@@ -5592,6 +5678,8 @@ var WebGPUEngine = class WebGPUEngine {
 	_outeTTS = null;
 	/** The preset voice the lazily-built OuteTTS engine was created with. */
 	_outeVoice = null;
+	/** Source of the runtime LoRA adapter currently applied on the static base, if any. */
+	_currentAdapter = null;
 	/** Lazily-created Parler-TTS engine (Flan-T5 encoder + decoder LM + dac_44khz). */
 	_parlerTTS = null;
 	/**
@@ -5601,6 +5689,20 @@ var WebGPUEngine = class WebGPUEngine {
 	* promotion runs at most once per session. Always false on Dawn/node.
 	*/
 	_groupProbePending = false;
+	/**
+	* Tail of the per-engine generation queue. The shared {@link Executor} owns a
+	* single set of readback/staging buffers (logits + argmax) and one KV cache, so
+	* two overlapping generate()/embed()/describeImage() calls on the SAME engine
+	* instance would call `mapAsync` on the logits/argmax readback while a prior map
+	* is still pending ("Buffer already has an outstanding map pending") and corrupt
+	* each other's KV state. The browser SDK shares one engine per model across
+	* components (see SHARED_ENGINES in browser/use-engine.ts), so this overlap is
+	* real — e.g. a playground's chat handle and its example-generator on the same
+	* instance. We serialize whole generations through this promise chain: a second
+	* caller awaits the first. Per-engine and only engaged on genuine overlap, so it
+	* adds no per-token overhead to the decode hot path.
+	*/
+	_genQueueTail = Promise.resolve();
 	/** Model capabilities (text, vision, moe). */
 	capabilities;
 	/** Model architecture config. */
@@ -5649,6 +5751,33 @@ var WebGPUEngine = class WebGPUEngine {
 		};
 	}
 	/**
+	* Load, swap, or remove a runtime LoRA adapter on the loaded base.
+	*
+	* @param source Adapter spec (`hf:owner/repo`, a URL, or `file:./dir` holding
+	*   `adapter_config.json` + `adapter_model.safetensors`), or `null` to drop it.
+	*/
+	async loadAdapter(source) {
+		if (source === null) {
+			this.executor.clearRuntimeLoRA();
+			this._currentAdapter = null;
+			return;
+		}
+		const adapter = await fetchAdapter(source, {
+			hfToken: this._createOptions.hfToken,
+			revision: this._createOptions.revision
+		});
+		if (!adapter) throw new Error(`No adapter found at ${source}.`);
+		const deltas = buildLoRADeltas(adapter, createKeyMapperForArch(this._architecture));
+		const { applied, skipped } = this.executor.applyRuntimeLoRA(deltas);
+		if (applied === 0) throw new Error(`Adapter ${source} resolved to 0 applicable runtime targets on ${this._architecture} (skipped ${skipped}).`);
+		console.log(`[engine] runtime adapter ${source}: applied ${applied} target(s)${skipped ? `, skipped ${skipped}` : ""}.`);
+		this._currentAdapter = source;
+	}
+	/** The runtime LoRA adapter currently applied on the base, or null. */
+	getAdapter() {
+		return this._currentAdapter;
+	}
+	/**
 	* Write a coarse crash-phase breadcrumb that survives a GPU-process kill / page
 	* reload. The iPad harness reads `localStorage["gerbil-crash-phase"]` after a
 	* crash; without these, a describe-time crash only shows the last load phase
@@ -5948,8 +6077,21 @@ var WebGPUEngine = class WebGPUEngine {
 	}
 	/**
 	* Generate text from a prompt.
+	*
+	* Serialized per engine: the shared executor's readback/staging buffers and KV
+	* cache cannot be shared by two overlapping generations, so a concurrent call on
+	* the same instance awaits this one (see {@link _genQueueTail}).
 	*/
 	async generate(prompt, options = {}) {
+		this.checkDestroyed();
+		const release = await this._acquireGenLock();
+		try {
+			return await this._generateLocked(prompt, options);
+		} finally {
+			release();
+		}
+	}
+	async _generateLocked(prompt, options = {}) {
 		this.checkDestroyed();
 		const { maxTokens = 512, stopSequences = [], sampling = {}, systemPrompt, onToken } = options;
 		this.executor.reset();
@@ -6017,7 +6159,7 @@ var WebGPUEngine = class WebGPUEngine {
 		let mmLogicalPos = inputIds.length;
 		const keepPlan = this.config.vocabKeepPlan;
 		const remapTok = keepPlan ? (i) => remapPrunedToken(i, keepPlan) : (i) => i;
-		if (isGreedy && !this.executor.needsMultiEncoder && !mmDecode) {
+		if (isGreedy && !this.executor.needsMultiEncoder && !mmDecode && !this.executor.hasRuntimeAdapter) {
 			const firstToken = remapTok(sampleToken(logits, sampling, [...inputIds, ...generatedIds]));
 			if (!consumeToken(firstToken)) {
 				const depth = Executor.PIPELINE_DEPTH;
@@ -6062,9 +6204,17 @@ var WebGPUEngine = class WebGPUEngine {
 	}
 	/**
 	* Inline autocomplete: continue `prefix` with a brief, single-line continuation.
-	* Wraps `generate` with low-latency defaults (16 tokens, temp 0.3, stop at the
-	* first newline) + a continuation system prompt, then cleans the output (strip
-	* after newline, dequote, drop an echoed prefix, smart leading space).
+	* Wraps `generate` with low-latency defaults (16 tokens, temp 0.3, a repetition
+	* penalty of 1.25, stop at the first newline) + a continuation system prompt,
+	* then cleans the output: strip after newline, dequote, drop an echoed prefix,
+	* drop a leading overlap with the typed tail, drop a leading word-run that
+	* restates an earlier typed phrase, truncate an internal phrase loop, cap the
+	* length, and add a smart leading space.
+	*
+	* The repetition penalty plus the overlap/restatement/loop cleanup keep small
+	* quantized models (e.g. Qwen3.5-0.8B) from echoing the typed text or looping
+	* on a phrase. Both are on by default; `repetitionPenalty` and `maxChars` are
+	* overridable.
 	*
 	* ```ts
 	* const suggestion = await engine.autocomplete("The quick brown fox");
@@ -6072,12 +6222,18 @@ var WebGPUEngine = class WebGPUEngine {
 	* ```
 	*/
 	async autocomplete(prefix, opts = {}) {
-		return normalizeContinuation((await this.generate(prefix, {
+		return cleanSuggestion((await this.generate(prefix, {
 			systemPrompt: AUTOCOMPLETE_SYSTEM,
 			maxTokens: opts.maxTokens ?? 16,
-			sampling: { temperature: opts.temperature ?? .3 },
+			sampling: {
+				temperature: opts.temperature ?? .3,
+				repetitionPenalty: opts.repetitionPenalty ?? 1.25
+			},
 			stopSequences: opts.stop ?? ["\n"]
-		})).text, prefix, opts.singleLine ?? true);
+		})).text, prefix, {
+			singleLine: opts.singleLine ?? true,
+			maxChars: opts.maxChars
+		});
 	}
 	/**
 	* Rewrite `text` in a target tone (e.g. "professional", "friendly", "concise",
@@ -6380,6 +6536,15 @@ var WebGPUEngine = class WebGPUEngine {
 	*    HF-exact pixel_values from a reference; skips host preprocessing).
 	*/
 	async describeImage(image, prompt = "Describe this image.", options = {}) {
+		this.checkDestroyed();
+		const release = await this._acquireGenLock();
+		try {
+			return await this._describeImageLocked(image, prompt, options);
+		} finally {
+			release();
+		}
+	}
+	async _describeImageLocked(image, prompt = "Describe this image.", options = {}) {
 		this.checkDestroyed();
 		if (!this.visionExecutor) throw new Error("describeImage() requires a vision encoder. Load with { enableVision: true } on a vision-capable checkpoint (Qwen3.5 or Gemma 4).");
 		if (this.visionExecutor.gemma4) {
@@ -6646,6 +6811,15 @@ var WebGPUEngine = class WebGPUEngine {
 	* `{ taskPrompt }` for other tasks (clustering/classification/STS).
 	*/
 	async embed(text, options = {}) {
+		this.checkDestroyed();
+		const release = await this._acquireGenLock();
+		try {
+			return await this._embedLocked(text, options);
+		} finally {
+			release();
+		}
+	}
+	async _embedLocked(text, options = {}) {
 		this.checkDestroyed();
 		if (!this._isEmbedding) throw new Error("embed() requires an embedding model. Load with { embedding: true } (e.g. repo: 'Qwen/Qwen3-Embedding-0.6B' or 'mlx-community/embeddinggemma-300m-4bit').");
 		this.executor.reset();
@@ -7072,8 +7246,30 @@ var WebGPUEngine = class WebGPUEngine {
 	checkDestroyed() {
 		if (this._destroyed) throw new Error("WebGPUEngine has been destroyed");
 	}
+	/**
+	* Acquire the per-engine generation lock (see {@link _genQueueTail}). Returns a
+	* release fn that MUST be called in a `finally` so the chain never wedges. This
+	* is a plain promise-chain mutex: each caller links onto the current tail and
+	* only proceeds once the previous holder releases.
+	*
+	* Non-reentrant: only the top-level generation entry points that directly drive
+	* the shared executor readback — generate(), embed(), describeImage() — acquire
+	* it. Helpers that delegate to generate() (autocomplete, rewrite,
+	* generateWithTools, generateObject, stream) inherit serialization through that
+	* single call and must NOT acquire it themselves, or they would self-deadlock.
+	*/
+	async _acquireGenLock() {
+		let release;
+		const next = new Promise((resolve) => {
+			release = resolve;
+		});
+		const prev = this._genQueueTail;
+		this._genQueueTail = next;
+		await prev;
+		return release;
+	}
 };
 
 //#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-kQVLpV3n.mjs.map
+//# sourceMappingURL=gpu-NPdk7pmr.mjs.map