@tryhamster/gerbil 1.6.1 → 1.6.3

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-KnrRSIeJ.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-CpmH4Gha.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-DtA28rES.mjs";
-import { a as loadMoonshine, d as Tokenizer, f as Executor, i as loadModel, l as quantizeKaniBackbone, n as MoonshineEncoderExecutor, r as loadKaniTTS, t as MoonshineSTT, x as initGPU } from "../moonshine-stt-BwX7NQ0O.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-BZVtzX9N.mjs";
+import { a as loadMoonshine, d as Tokenizer, f as Executor, i as loadModel, l as quantizeKaniBackbone, n as MoonshineEncoderExecutor, r as loadKaniTTS, t as MoonshineSTT, x as initGPU } from "../moonshine-stt-BFIm_Xxk.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, audioTokensToCodes, audioTokensToDacCodes, buildGemma4PoolMatrix, buildGemma4PosEmbeds, buildGemma4RotaryCosSin, buildGemma4VisionPositionTensors, buildMRoPECosSin, buildMRoPEPositionIds, buildOutePromptString, buildPosEmbeds, buildPositionIds, buildRotaryCosSin, buildVisionPositionTensors, dacOutputLength, dequantizeGemma4VisionProjection, dequantizeMLXProjection, 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-DtA28rES.mjs → gpu-BZVtzX9N.mjs}

@@ -1,6 +1,6 @@
 import { i as resolveDefaultRepo, n as OUTETTS_ASSETS, r as OUTETTS_PRESET_VOICES, t as DEFAULT_MODELS } from "./defaults-KnrRSIeJ.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-CpmH4Gha.mjs";
-import { S as verifyGPU, _ as createStorageBuffer, b as getOrCreatePipeline, c as quantizeBackboneInt4, f as Executor, g as createBindGroup, h as clearPipelineCache, i as loadModel, l as quantizeKaniBackbone, m as MATMUL_BIAS_F16C_SPEC, o as loadOuteTTS, p as KERNEL_REGISTRY, r as loadKaniTTS, s as loadParlerTTS, u as remapPrunedToken, v as createUniformBuffer, x as initGPU, y as destroyBuffers } from "./moonshine-stt-BwX7NQ0O.mjs";
+import { S as verifyGPU, _ as createStorageBuffer, b as getOrCreatePipeline, c as quantizeBackboneInt4, f as Executor, g as createBindGroup, h as clearPipelineCache, i as loadModel, l as quantizeKaniBackbone, m as MATMUL_BIAS_F16C_SPEC, o as loadOuteTTS, p as KERNEL_REGISTRY, r as loadKaniTTS, s as loadParlerTTS, u as remapPrunedToken, v as createUniformBuffer, x as initGPU, y as destroyBuffers } from "./moonshine-stt-BFIm_Xxk.mjs";
 
 //#region src/gpu/architectures/gemma4_vision.ts
 /**
@@ -4424,11 +4424,24 @@ var VisionExecutor = class {
 			q.writeBuffer(d.uniformBuffer, 0, params);
 			sizes.push(d.spec.getDispatchSize(d.node, resolvedShapes, ctxRt));
 		}
-		if (this.ctx.isWebKitWebGPU) {
+		const multiEncoder = this.ctx.isWebKitWebGPU;
+		const batchVit = globalThis.GERBIL_VIT_BATCH != null;
+		let batchEnc = null;
+		let batchPass = null;
+		const flushBatch = async () => {
+			if (!batchEnc || !batchPass) return;
+			batchPass.end();
+			q.submit([batchEnc.finish()]);
+			await q.onSubmittedWorkDone();
+			batchEnc = null;
+			batchPass = null;
+		};
+		if (multiEncoder) {
 			onStage?.("vit-encode-start", { total: this.dispatches.length });
 			for (let i = 0; i < this.dispatches.length; i++) {
 				const d = this.dispatches[i];
 				if (d.node.opType === "Attention") {
+					await flushBatch();
 					const [gridX, gridY, gridZ] = sizes[i];
 					const layer = Number.parseInt(d.node.id.replace(/\D+/g, ""), 10);
 					if (!Number.isNaN(layer)) onStage?.("vit-attn", { layer });
@@ -4450,6 +4463,16 @@ var VisionExecutor = class {
 					}
 					continue;
 				}
+				if (batchVit) {
+					if (!batchEnc) {
+						batchEnc = this.ctx.device.createCommandEncoder();
+						batchPass = batchEnc.beginComputePass();
+					}
+					batchPass.setPipeline(d.pipeline);
+					batchPass.setBindGroup(0, d.bindGroup);
+					batchPass.dispatchWorkgroups(...sizes[i]);
+					continue;
+				}
 				const enc$1 = this.ctx.device.createCommandEncoder();
 				const p = enc$1.beginComputePass();
 				p.setPipeline(d.pipeline);
@@ -4459,6 +4482,7 @@ var VisionExecutor = class {
 				q.submit([enc$1.finish()]);
 				await q.onSubmittedWorkDone();
 			}
+			await flushBatch();
 			onStage?.("vit-encode-done");
 		} else {
 			const enc$1 = this.ctx.device.createCommandEncoder({ label: "vision_encode" });
@@ -4634,8 +4658,24 @@ var VisionExecutor = class {
 		return resolved;
 	}
 	allocateActivationBuffers() {
+		const referenced = /* @__PURE__ */ new Set();
+		for (const node of this.graph.nodes) {
+			for (const t of node.inputs) referenced.add(t);
+			for (const t of node.outputs) referenced.add(t);
+		}
+		let skippedBytes = 0;
 		for (const [name, desc] of Object.entries(this.graph.tensors)) {
 			if (desc.storage !== "activation") continue;
+			if (!referenced.has(name)) {
+				const shape = desc.shape.map((d) => {
+					if (d === "N") return this.maxPatches;
+					if (d === "Nm") return this.maxPatches / this.mergeUnit;
+					if (d === "Np") return this.maxPooled();
+					return d;
+				});
+				skippedBytes += shape.reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];
+				continue;
+			}
 			const bytes = desc.shape.map((d) => {
 				if (d === "N") return this.maxPatches;
 				if (d === "Nm") return this.maxPatches / this.mergeUnit;
@@ -4644,6 +4684,7 @@ var VisionExecutor = class {
 			}).reduce((a, b) => a * b, 1) * DTYPE_BYTES[desc.dtype];
 			this.activationBuffers.set(name, createStorageBuffer(this.ctx, `vact_${name}`, bytes));
 		}
+		if (skippedBytes > 0) console.log(`[vision] skipped ${(skippedBytes / 1e6).toFixed(0)} MB of unreferenced activation tensors`);
 	}
 	gatherBuffers(spec, node, uniformBuffer) {
 		const entries = [];
@@ -7035,4 +7076,4 @@ 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-DtA28rES.mjs.map
+//# sourceMappingURL=gpu-BZVtzX9N.mjs.map