@tryhamster/gerbil 1.2.0 → 1.2.1

package/dist/gpu/hooks.d.mts

@@ -1,4 +1,4 @@
-import { n as AgentTool, t as AgentStep } from "../index-CSLqgNu-.mjs";
+import { n as AgentTool, t as AgentStep } from "../index-HVS7yeLx.mjs";
 import { c as MemorySearchResult, d as RecallOptions, f as RecallResult, m as SearchOptions, s as MemoryRecord, t as AddOptions } from "../types-DQBe2lFo.mjs";
 import { ReactNode } from "react";
 import * as react_jsx_runtime0 from "react/jsx-runtime";

package/dist/gpu/index.d.mts

@@ -1,2 +1,2 @@
-import { $ as buildMRoPECosSin, A as generateKaniTtsGraph, B as resolveDefaultRepo, C as generateQwen3_5VisionGraph, Ct as initGPU, D as moonshineEncoderFrames, Dt as ModelArchConfig, E as generateMoonshineEncoderGraph, Et as KvMode, F as dequantizeMLXProjection, G as PreprocessedImage, H as Gemma4VisionGridConfig, I as generateGemma4VisionGraph, J as VisionPositionTensors, K as QWEN3_5_IMAGE_PROCESSOR, L as patchGemma4VisionClips, M as parseKaniConfig, N as Gemma4VisionGraphInfo, O as parseMoonshineConfig, Ot as ModelCapabilities, P as dequantizeGemma4VisionProjection, Q as buildGemma4VisionPositionTensors, R as resolveGemma4VisionInfo, S as Executor, St as GPUDiagnosticResult, T as generateMoonshineDecoderGraph, Tt as KVDType, U as Gemma4VisionPositionTensors, V as GEMMA4_IMAGE_PROCESSOR, W as ImageProcessorConfig, X as buildGemma4PosEmbeds, Y as buildGemma4PoolMatrix, Z as buildGemma4RotaryCosSin, _ as MoonshineSTT, _t as KaniDType, a as GenerateObjectOptions, at as mropeFreqDims, b as TranscribeResult, bt as SpeakOptions, c as GenerateResult, ct as smartResize, d as ObjectSchema, dt as LoadedMoonshine, et as buildMRoPEPositionIds, f as ObjectValidator, ft as loadKaniTTS, g as VisionInputs, gt as ChatMessage, h as VisionExecutor, ht as quantizeKaniBackbone, i as EncodeImageResult, it as buildVisionPositionTensors, j as generateNanoCodecDecoderGraph, k as audioTokensToCodes, l as IntegrityCheckEntry, lt as SamplingParams, m as WebGPUEngineOptions, mt as loadMoonshine, n as AgentTool, nt as buildPositionIds, o as GenerateObjectResult, ot as preprocessImage, p as WebGPUEngine, pt as loadModel, q as VisionGridConfig, r as EmbedOptions, rt as buildRotaryCosSin, s as GenerateOptions, st as preprocessImageGemma4, t as AgentStep, tt as buildPosEmbeds, u as IntegrityCheckResult, ut as LoadedKaniTTS, v as MoonshineSTTOptions, vt as KaniTTS, w as MOONSHINE_REMAINING_WORK, wt as GraphDType, x as MoonshineEncoderExecutor, xt as SpeakResult, y as TranscribeOptions, yt as KaniTTSOptions, z as DEFAULT_MODELS } from "../index-CSLqgNu-.mjs";
+import { $ as buildMRoPECosSin, A as generateKaniTtsGraph, B as resolveDefaultRepo, C as generateQwen3_5VisionGraph, Ct as initGPU, D as moonshineEncoderFrames, Dt as ModelArchConfig, E as generateMoonshineEncoderGraph, Et as KvMode, F as dequantizeMLXProjection, G as PreprocessedImage, H as Gemma4VisionGridConfig, I as generateGemma4VisionGraph, J as VisionPositionTensors, K as QWEN3_5_IMAGE_PROCESSOR, L as patchGemma4VisionClips, M as parseKaniConfig, N as Gemma4VisionGraphInfo, O as parseMoonshineConfig, Ot as ModelCapabilities, P as dequantizeGemma4VisionProjection, Q as buildGemma4VisionPositionTensors, R as resolveGemma4VisionInfo, S as Executor, St as GPUDiagnosticResult, T as generateMoonshineDecoderGraph, Tt as KVDType, U as Gemma4VisionPositionTensors, V as GEMMA4_IMAGE_PROCESSOR, W as ImageProcessorConfig, X as buildGemma4PosEmbeds, Y as buildGemma4PoolMatrix, Z as buildGemma4RotaryCosSin, _ as MoonshineSTT, _t as KaniDType, a as GenerateObjectOptions, at as mropeFreqDims, b as TranscribeResult, bt as SpeakOptions, c as GenerateResult, ct as smartResize, d as ObjectSchema, dt as LoadedMoonshine, et as buildMRoPEPositionIds, f as ObjectValidator, ft as loadKaniTTS, g as VisionInputs, gt as ChatMessage, h as VisionExecutor, ht as quantizeKaniBackbone, i as EncodeImageResult, it as buildVisionPositionTensors, j as generateNanoCodecDecoderGraph, k as audioTokensToCodes, l as IntegrityCheckEntry, lt as SamplingParams, m as WebGPUEngineOptions, mt as loadMoonshine, n as AgentTool, nt as buildPositionIds, o as GenerateObjectResult, ot as preprocessImage, p as WebGPUEngine, pt as loadModel, q as VisionGridConfig, r as EmbedOptions, rt as buildRotaryCosSin, s as GenerateOptions, st as preprocessImageGemma4, t as AgentStep, tt as buildPosEmbeds, u as IntegrityCheckResult, ut as LoadedKaniTTS, v as MoonshineSTTOptions, vt as KaniTTS, w as MOONSHINE_REMAINING_WORK, wt as GraphDType, x as MoonshineEncoderExecutor, xt as SpeakResult, y as TranscribeOptions, yt as KaniTTSOptions, z as DEFAULT_MODELS } from "../index-HVS7yeLx.mjs";
 export { 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, LoadedKaniTTS, LoadedMoonshine, MOONSHINE_REMAINING_WORK, ModelArchConfig, ModelCapabilities, MoonshineEncoderExecutor, MoonshineSTT, MoonshineSTTOptions, ObjectSchema, ObjectValidator, PreprocessedImage, QWEN3_5_IMAGE_PROCESSOR, SamplingParams, SpeakOptions, SpeakResult, TranscribeOptions, TranscribeResult, VisionExecutor, VisionGridConfig, VisionInputs, VisionPositionTensors, WebGPUEngine, WebGPUEngineOptions, audioTokensToCodes, buildGemma4PoolMatrix, buildGemma4PosEmbeds, buildGemma4RotaryCosSin, buildGemma4VisionPositionTensors, buildMRoPECosSin, buildMRoPEPositionIds, buildPosEmbeds, buildPositionIds, buildRotaryCosSin, buildVisionPositionTensors, dequantizeGemma4VisionProjection, dequantizeMLXProjection, generateGemma4VisionGraph, generateKaniTtsGraph, generateMoonshineDecoderGraph, generateMoonshineEncoderGraph, generateNanoCodecDecoderGraph, generateQwen3_5VisionGraph, initGPU, loadKaniTTS, loadModel, loadMoonshine, moonshineEncoderFrames, mropeFreqDims, parseKaniConfig, parseMoonshineConfig, patchGemma4VisionClips, preprocessImage, preprocessImageGemma4, quantizeKaniBackbone, resolveDefaultRepo, resolveGemma4VisionInfo, smartResize };

package/dist/gpu/index.mjs

@@ -1,6 +1,6 @@
 import { n as resolveDefaultRepo, t as DEFAULT_MODELS } from "../defaults-bJcmOaRG.mjs";
 import { 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-Ol7yUmKf.mjs";
-import { C as generateGemma4VisionGraph, S as dequantizeMLXProjection, T as resolveGemma4VisionInfo, _ as smartResize, a as buildGemma4PosEmbeds, b as generateQwen3_5VisionGraph, c as buildMRoPECosSin, d as buildPositionIds, f as buildRotaryCosSin, g as preprocessImageGemma4, h as preprocessImage, i as buildGemma4PoolMatrix, 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 patchGemma4VisionClips, x as dequantizeGemma4VisionProjection, y as KaniTTS } from "../gpu-BWLMlkVW.mjs";
-import { _ as initGPU, a as loadMoonshine, c as Executor, i as loadModel, n as MoonshineEncoderExecutor, o as quantizeKaniBackbone, r as loadKaniTTS, t as MoonshineSTT } from "../moonshine-stt-C4GKK-h5.mjs";
+import { C as generateGemma4VisionGraph, S as dequantizeMLXProjection, T as resolveGemma4VisionInfo, _ as smartResize, a as buildGemma4PosEmbeds, b as generateQwen3_5VisionGraph, c as buildMRoPECosSin, d as buildPositionIds, f as buildRotaryCosSin, g as preprocessImageGemma4, h as preprocessImage, i as buildGemma4PoolMatrix, 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 patchGemma4VisionClips, x as dequantizeGemma4VisionProjection, y as KaniTTS } from "../gpu-DR6EG0K-.mjs";
+import { _ as initGPU, a as loadMoonshine, c as Executor, i as loadModel, n as MoonshineEncoderExecutor, o as quantizeKaniBackbone, r as loadKaniTTS, t as MoonshineSTT } from "../moonshine-stt-DSz3VW_0.mjs";
 
 export { DEFAULT_MODELS, Executor, GEMMA4_IMAGE_PROCESSOR, KaniTTS, MOONSHINE_REMAINING_WORK, MoonshineEncoderExecutor, MoonshineSTT, QWEN3_5_IMAGE_PROCESSOR, VisionExecutor, WebGPUEngine, audioTokensToCodes, buildGemma4PoolMatrix, buildGemma4PosEmbeds, buildGemma4RotaryCosSin, buildGemma4VisionPositionTensors, buildMRoPECosSin, buildMRoPEPositionIds, buildPosEmbeds, buildPositionIds, buildRotaryCosSin, buildVisionPositionTensors, dequantizeGemma4VisionProjection, dequantizeMLXProjection, generateGemma4VisionGraph, generateKaniTtsGraph, generateMoonshineDecoderGraph, generateMoonshineEncoderGraph, generateNanoCodecDecoderGraph, generateQwen3_5VisionGraph, initGPU, loadKaniTTS, loadModel, loadMoonshine, moonshineEncoderFrames, mropeFreqDims, parseKaniConfig, parseMoonshineConfig, patchGemma4VisionClips, preprocessImage, preprocessImageGemma4, quantizeKaniBackbone, resolveDefaultRepo, resolveGemma4VisionInfo, smartResize };

package/dist/{gpu-BWLMlkVW.mjs → gpu-DR6EG0K-.mjs}

@@ -1,6 +1,6 @@
 import { n as resolveDefaultRepo, t as DEFAULT_MODELS } from "./defaults-bJcmOaRG.mjs";
 import { E as GEMMA4_VIS_KEYS, 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-Ol7yUmKf.mjs";
-import { _ as initGPU, c as Executor, d as clearPipelineCache, f as createBindGroup, g as getOrCreatePipeline, h as destroyBuffers, i as loadModel, l as KERNEL_REGISTRY, m as createUniformBuffer, o as quantizeKaniBackbone, p as createStorageBuffer, r as loadKaniTTS, s as remapPrunedToken, u as MATMUL_BIAS_F16C_SPEC, v as verifyGPU } from "./moonshine-stt-C4GKK-h5.mjs";
+import { _ as initGPU, c as Executor, d as clearPipelineCache, f as createBindGroup, g as getOrCreatePipeline, h as destroyBuffers, i as loadModel, l as KERNEL_REGISTRY, m as createUniformBuffer, o as quantizeKaniBackbone, p as createStorageBuffer, r as loadKaniTTS, s as remapPrunedToken, u as MATMUL_BIAS_F16C_SPEC, v as verifyGPU } from "./moonshine-stt-DSz3VW_0.mjs";
 
 //#region src/gpu/architectures/gemma4_vision.ts
 /**
@@ -855,7 +855,7 @@ var KaniTTS = class KaniTTS {
 		this.rawConfig = loaded.rawConfig;
 		this.cfg = parseKaniConfig(loaded.rawConfig);
 		this.maxSeqLen = maxSeqLen;
-		this.dtype = dtype;
+		this.dtype = loaded.preQuantized ? "q4" : dtype;
 		this.architecture = (loaded.rawConfig.architectures ?? [])[0] ?? "Lfm2ForCausalLM";
 		const langSettings = loaded.rawConfig.language_settings;
 		this.voiceTags = Array.isArray(langSettings?.language_tags_list) ? langSettings.language_tags_list.filter((v) => typeof v === "string") : [];
@@ -874,14 +874,26 @@ var KaniTTS = class KaniTTS {
 			}
 			this.layerAlpha.set(layer, alpha);
 		}
-		const graph = generateKaniTtsGraph(this.rawConfig, dtype === "q4" ? "q4" : void 0);
-		if (dtype === "q4") quantizeKaniBackbone(graph, loaded.backboneWeights);
-		this.backboneExec = new Executor(ctx, graph, {
+		this.backboneExec = this.buildBackbone(ctx, loaded, maxSeqLen);
+	}
+	/**
+	* Build the codec-LM backbone graph, quantize it when q4, and upload its weights
+	* into a fresh Executor. For a pre-quantized (gerbil-int4) checkpoint the linear
+	* projections already carry `.q`/`.scales`/`.zeros`, so quantizeKaniBackbone is an
+	* idempotent no-op (it skips any projection whose `.q` is already present) and the
+	* int4 weights are uploaded verbatim — no on-device re-quantization.
+	*/
+	buildBackbone(ctx, loaded, maxSeqLen) {
+		const useQ4 = this.dtype === "q4";
+		const graph = generateKaniTtsGraph(this.rawConfig, useQ4 ? "q4" : void 0);
+		if (useQ4) quantizeKaniBackbone(graph, loaded.backboneWeights);
+		const exec = new Executor(ctx, graph, {
 			maxSeqLen,
 			kvMode: "f32"
 		});
-		this.backboneExec.uploadWeightsMap(selectGraphWeights(graph, loaded.backboneWeights));
-		this.backboneExec.initBindGroups();
+		exec.uploadWeightsMap(selectGraphWeights(graph, loaded.backboneWeights));
+		exec.initBindGroups();
+		return exec;
 	}
 	static async create(options = {}) {
 		const ctx = await initGPU();
@@ -3856,4 +3868,4 @@ var WebGPUEngine = class WebGPUEngine {
 
 //#endregion
 export { generateGemma4VisionGraph as C, dequantizeMLXProjection as S, resolveGemma4VisionInfo as T, smartResize as _, buildGemma4PosEmbeds as a, generateQwen3_5VisionGraph as b, buildMRoPECosSin as c, buildPositionIds as d, buildRotaryCosSin as f, preprocessImageGemma4 as g, preprocessImage as h, buildGemma4PoolMatrix as i, 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, patchGemma4VisionClips as w, dequantizeGemma4VisionProjection as x, KaniTTS as y };
-//# sourceMappingURL=gpu-BWLMlkVW.mjs.map
+//# sourceMappingURL=gpu-DR6EG0K-.mjs.map