@tryhamster/gerbil 1.6.0 → 1.6.2

package/dist/{gpu-DfUBMUGe.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-Ol7yUmKf.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-bExSnd2o.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-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 = [];
@@ -5324,12 +5365,16 @@ function writeGroupProbe(rec) {
 	} catch {}
 }
 /**
-* The validated non-phone sweet spot. iPad swept 1→7.9, 8→19, 32→24.8, 64→26.6,
-* 128→26.9 (peak), 256→26.2 tok/s — a plateau from ~64 up, so 128 is the best
-* stable target (more batching just costs memory). Non-phone WebKit jumps here
-* directly; the crash breadcrumb caps it down if a device can't sustain it.
+* The validated batching target. Originally 128 from an iPad sweep on a much
+* slower engine build (peak 26.9 tok/s at 128, 256 within 2.6%); the 2026-07
+* kernel campaign made per-dispatch GPU time far smaller, which moves the
+* amortization optimum UP. An on-device iPhone sweep (bench harness, engine
+* 1.6.1) measured Qwen3.5-0.8B sampled decode: g32=19, g128≈33, g256=43,
+* g1024=45, g4096=44 — a +36% win at 1024 vs the old 128 target — while
+* LFM2.5 is flat from g32 upward (no cost). The crash breadcrumb still caps
+* any device that can't sustain it.
 */
-const NONPHONE_TARGET_GROUP = 128;
+const NONPHONE_TARGET_GROUP = 1024;
 /**
 * Resolve the WebKit group size to use this session, recording `trying` as a
 * side effect so a crash this load is detectable on the next load.
@@ -7031,4 +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-DfUBMUGe.mjs.map
+//# sourceMappingURL=gpu-BZVtzX9N.mjs.map