@tryhamster/gerbil 1.11.2 → 1.11.4

package/dist/{gpu-BQ8I-VDx.mjs → gpu-CrzjQHv2.mjs}

@@ -1,6 +1,6 @@
 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 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";
+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-BXoZaHJE.mjs";
 
 //#region src/gpu/architectures/gemma4_vision.ts
 /**
@@ -5579,6 +5579,19 @@ var WebGPUEngine = class WebGPUEngine {
 			release();
 		}
 	}
+	/**
+	* Resolve the end-of-turn stop token id, or null if the model has none.
+	*
+	* Chat models like Gemma 4 end an assistant turn with a dedicated end-of-turn
+	* token (`<turn|>`, id 106) rather than the generic `<eos>`. Every text-decode
+	* loop must treat it as a stop token or the model runs on past its answer and
+	* repeats until it hits `maxTokens`. Resolved from the tokenizer's added tokens
+	* so it is null for models without a turn delimiter (Qwen, LFM2, …), leaving
+	* their decode behaviour unchanged.
+	*/
+	resolveEndOfTurnId() {
+		return this.tokenizer.tokenToId("<turn|>") ?? this.tokenizer.tokenToId("<end_of_turn>");
+	}
 	async _generateLocked(prompt, options = {}) {
 		this.checkDestroyed();
 		const { maxTokens = 512, stopSequences = [], sampling = {}, systemPrompt, onToken } = options;
@@ -5615,10 +5628,11 @@ var WebGPUEngine = class WebGPUEngine {
 		let finishReason = "max_tokens";
 		let generatedText = "";
 		const eosId = this.tokenizer.config.eosTokenId;
+		const eotId = this.resolveEndOfTurnId();
 		const decodeStart = performance.now();
 		const consumeToken = (nextToken) => {
 			generatedIds.push(nextToken);
-			if (eosId !== null && nextToken === eosId) {
+			if (eosId !== null && nextToken === eosId || eotId !== null && nextToken === eotId) {
 				finishReason = "eos";
 				return true;
 			}
@@ -5645,9 +5659,10 @@ var WebGPUEngine = class WebGPUEngine {
 		};
 		const mmDecode = hasMRoPE;
 		let mmLogicalPos = inputIds.length;
+		const streamsPle = this.executor.hasPleSource();
 		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 && !streamsPle) {
 			const firstToken = remapTok(sampleToken(logits, sampling, [...inputIds, ...generatedIds]));
 			if (!consumeToken(firstToken)) {
 				const depth = Executor.PIPELINE_DEPTH;
@@ -5664,19 +5679,22 @@ var WebGPUEngine = class WebGPUEngine {
 					if (consumeToken(tok)) break;
 				}
 			}
-		} else for (let step = 0; step < maxTokens; step++) {
-			let nextToken;
-			if (step === 0 || !isGreedy) nextToken = remapTok(sampleToken(logits, sampling, [...inputIds, ...generatedIds]));
-			else {
-				if (mmDecode) this.writeMRoPEDecodeStep(this.executor.currentSeqPos, mmLogicalPos);
-				nextToken = await this.executor.forwardArgmax(new Uint32Array([generatedIds[generatedIds.length - 1]]));
-				mmLogicalPos++;
-			}
-			if (consumeToken(nextToken)) break;
-			if (!isGreedy) {
-				if (mmDecode) this.writeMRoPEDecodeStep(this.executor.currentSeqPos, mmLogicalPos);
-				logits = (await this.executor.forward(new Uint32Array([nextToken]))).logits;
-				mmLogicalPos++;
+		} else {
+			const useCpuForward = !isGreedy || streamsPle;
+			for (let step = 0; step < maxTokens; step++) {
+				let nextToken;
+				if (step === 0 || useCpuForward) nextToken = remapTok(sampleToken(logits, sampling, [...inputIds, ...generatedIds]));
+				else {
+					if (mmDecode) this.writeMRoPEDecodeStep(this.executor.currentSeqPos, mmLogicalPos);
+					nextToken = await this.executor.forwardArgmax(new Uint32Array([generatedIds[generatedIds.length - 1]]));
+					mmLogicalPos++;
+				}
+				if (consumeToken(nextToken)) break;
+				if (useCpuForward) {
+					if (mmDecode) this.writeMRoPEDecodeStep(this.executor.currentSeqPos, mmLogicalPos);
+					logits = (await this.executor.forward(new Uint32Array([nextToken]))).logits;
+					mmLogicalPos++;
+				}
 			}
 		}
 		const totalTime = performance.now() - startTime;
@@ -6161,7 +6179,7 @@ var WebGPUEngine = class WebGPUEngine {
 		let finishReason = "max_tokens";
 		let generatedText = "";
 		const eosId = this.tokenizer.config.eosTokenId;
-		const eotId = this.tokenizer.tokenToId("<turn|>");
+		const eotId = this.resolveEndOfTurnId();
 		const consumeToken = (nextToken) => {
 			generatedIds.push(nextToken);
 			if (eosId !== null && nextToken === eosId || eotId !== null && nextToken === eotId) {
@@ -6238,10 +6256,11 @@ var WebGPUEngine = class WebGPUEngine {
 		let finishReason = "max_tokens";
 		let generatedText = "";
 		const eosId = this.tokenizer.config.eosTokenId;
+		const eotId = this.resolveEndOfTurnId();
 		let mmLogicalPos = decodeStartPos;
 		const consumeToken = (nextToken) => {
 			generatedIds.push(nextToken);
-			if (eosId !== null && nextToken === eosId) {
+			if (eosId !== null && nextToken === eosId || eotId !== null && nextToken === eotId) {
 				finishReason = "eos";
 				return true;
 			}
@@ -6762,4 +6781,4 @@ 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-BQ8I-VDx.mjs.map
+//# sourceMappingURL=gpu-CrzjQHv2.mjs.map