@tryhamster/gerbil 1.10.1 → 1.10.2

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-B1xk3xOJ.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-Ps1eQaXG.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-BKeD6OYF.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-B1xk3xOJ.mjs → gpu-Ps1eQaXG.mjs}

@@ -806,7 +806,7 @@ function nucleusSample$1(probs, ids, topP) {
 * In-place top-k filter over a probability array: keep only the `k` highest, zero
 * the rest, and renormalize. No-op when k <= 0 or k >= length (keep everything).
 */
-function applyTopK$1(scores, k) {
+function applyTopK(scores, k) {
 	const n = scores.length;
 	if (k <= 0 || k >= n) return;
 	const order = Array.from({ length: n }, (_, i) => i).sort((a, b) => scores[b] - scores[a]);
@@ -1065,7 +1065,7 @@ var KaniTTS = class KaniTTS {
 			scores[i] = s / p.temperature;
 		}
 		softmaxInPlace$1(scores);
-		applyTopK$1(scores, p.topK);
+		applyTopK(scores, p.topK);
 		return nucleusSample$1(scores, ids, p.topP);
 	}
 	/**
@@ -2880,16 +2880,6 @@ function softmaxInPlace(scores) {
 	}
 	for (let i = 0; i < scores.length; i++) scores[i] /= sum;
 }
-/** In-place top-k filter: keep the k highest, zero the rest, renormalize. */
-function applyTopK(scores, k) {
-	const n = scores.length;
-	if (k <= 0 || k >= n) return;
-	const order = Array.from({ length: n }, (_, i) => i).sort((a, b) => scores[b] - scores[a]);
-	let keepSum = 0;
-	for (let r = 0; r < k; r++) keepSum += scores[order[r]];
-	for (let r = k; r < n; r++) scores[order[r]] = 0;
-	if (keepSum > 0) for (let i = 0; i < n; i++) scores[i] /= keepSum;
-}
 /** Top-p nucleus sample from a softmaxed prob array over `ids`. */
 function nucleusSample(probs, ids, topP) {
 	const n = probs.length;
@@ -2913,6 +2903,81 @@ function nucleusSample(probs, ids, topP) {
 	return ids[order[0]];
 }
 /**
+* Select the `k` highest-scoring token ids in a single O(n·logk) pass (bounded
+* min-heap), applying the repetition penalty to `previous` tokens inline so a
+* penalized token can drop out of the shortlist. Returns the kept ids and their
+* (penalized, pre-temperature) logits, unsorted.
+*
+* This replaces the two full-vocab `Array.from(...).sort(...)` passes the sampler
+* ran PER TOKEN over OuteTTS's ~157k vocab (the top-k filter + the nucleus sort):
+* a codec-LM emits hundreds-to-thousands of tokens per utterance, so those two
+* O(n·logn) sorts — plus their two fresh 157k-element allocations every step —
+* dominated host-side decode and made the browser SPEECH tab look stalled
+* ("SYNTHESIZING…" for minutes). Restricting the softmax + nucleus to the kept
+* top-k is distribution-identical to the old full-vocab path (softmax is
+* monotonic and the discarded tail is renormalized away either way).
+*/
+function selectTopKPenalized(logits, k, penalized, penalty) {
+	const n = logits.length;
+	const cap = Math.min(k, n);
+	const heapVal = new Float32Array(cap);
+	const heapId = new Int32Array(cap);
+	let size = 0;
+	const penalize = (i, v) => {
+		if (penalized?.has(i)) return v > 0 ? v / penalty : v * penalty;
+		return v;
+	};
+	const siftDown$1 = () => {
+		let root = 0;
+		for (;;) {
+			let smallest = root;
+			const l = 2 * root + 1;
+			const r = 2 * root + 2;
+			if (l < size && heapVal[l] < heapVal[smallest]) smallest = l;
+			if (r < size && heapVal[r] < heapVal[smallest]) smallest = r;
+			if (smallest === root) break;
+			const tv = heapVal[root];
+			const ti = heapId[root];
+			heapVal[root] = heapVal[smallest];
+			heapId[root] = heapId[smallest];
+			heapVal[smallest] = tv;
+			heapId[smallest] = ti;
+			root = smallest;
+		}
+	};
+	const siftUp = (node0) => {
+		let node = node0;
+		while (node > 0) {
+			const parent = node - 1 >> 1;
+			if (heapVal[parent] <= heapVal[node]) break;
+			const tv = heapVal[node];
+			const ti = heapId[node];
+			heapVal[node] = heapVal[parent];
+			heapId[node] = heapId[parent];
+			heapVal[parent] = tv;
+			heapId[parent] = ti;
+			node = parent;
+		}
+	};
+	for (let i = 0; i < n; i++) {
+		const v = penalize(i, logits[i]);
+		if (size < cap) {
+			heapVal[size] = v;
+			heapId[size] = i;
+			size++;
+			siftUp(size - 1);
+		} else if (v > heapVal[0]) {
+			heapVal[0] = v;
+			heapId[0] = i;
+			siftDown$1();
+		}
+	}
+	return {
+		ids: heapId.subarray(0, size),
+		vals: heapVal.subarray(0, size)
+	};
+}
+/**
 * Resample interleaved-or-mono PCM to mono at `targetRate` via linear interpolation.
 * Audio is assumed already mono (Float32 [-1,1]); if `srcRate === targetRate` the
 * input is returned as-is. Linear resampling is sufficient for codec conditioning
@@ -3145,9 +3210,16 @@ var OuteTTS = class OuteTTS {
 	*/
 	sampleToken(logits, p) {
 		const n = logits.length;
+		const prev = p.repetitionPenalty !== 1 ? new Set(p.previous) : null;
+		if (p.topK > 0 && p.topK < n) {
+			const { ids: kIds, vals: kVals } = selectTopKPenalized(logits, p.topK, prev, p.repetitionPenalty);
+			const scores$1 = new Float32Array(kVals.length);
+			for (let i = 0; i < kVals.length; i++) scores$1[i] = kVals[i] / p.temperature;
+			softmaxInPlace(scores$1);
+			return nucleusSample(scores$1, Int32Array.from(kIds), p.topP);
+		}
 		const ids = new Int32Array(n);
 		const scores = new Float32Array(n);
-		const prev = p.repetitionPenalty !== 1 ? new Set(p.previous) : null;
 		for (let i = 0; i < n; i++) {
 			ids[i] = i;
 			let s = logits[i];
@@ -3155,7 +3227,6 @@ var OuteTTS = class OuteTTS {
 			scores[i] = s / p.temperature;
 		}
 		softmaxInPlace(scores);
-		applyTopK(scores, p.topK);
 		return nucleusSample(scores, ids, p.topP);
 	}
 	/**
@@ -7272,4 +7343,4 @@ 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-B1xk3xOJ.mjs.map
+//# sourceMappingURL=gpu-Ps1eQaXG.mjs.map