@tryhamster/gerbil 1.11.1 → 1.11.3

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 GEMMA4_IMAGE_PROCESSOR, A as generateDacSpeechDecoderGraph, At as AdapterConfig, B as generateNanoCodecDecoderGraph, Bt as SpeakResult, 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 KvMode, H as Gemma4VisionGraphInfo, Ht as initGPU, I as moonshineEncoderFrames, It as KaniDType, J as DEFAULT_MODELS, Jt as createDefaultHFKeyMapper, K as patchGemma4VisionClips, Kt as ModelArchConfig, L as parseMoonshineConfig, Lt as KaniTTS, 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 KaniTTSOptions, S as MoonshineSTTOptions, St as LoadedKaniTTS, T as MoonshineEncoderExecutor, Tt as loadModel, U as dequantizeGemma4VisionProjection, Ut as GraphDType, V as parseKaniConfig, Vt as GPUDiagnosticResult, W as dequantizeMLXProjection, Wt as KVDType, X as OUTETTS_PRESET_VOICES, Y as OUTETTS_ASSETS, 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 ModelCapabilities, 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 SpeakOptions } from "../index-DqJMofyU.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 };
+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-DeKYzJgJ.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 { 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-BCfd_K38.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-EyIH7Qfc.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-BXoZaHJE.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, isTTSRepo, 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-BCfd_K38.mjs → gpu-EyIH7Qfc.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 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-BXoZaHJE.mjs";
 
 //#region src/gpu/architectures/gemma4_vision.ts
 /**
@@ -5645,9 +5645,10 @@ 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 +5665,22 @@ 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;
@@ -6762,4 +6766,4 @@ 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-BCfd_K38.mjs.map
+//# sourceMappingURL=gpu-EyIH7Qfc.mjs.map