runntime 0.0.1-alpha.1 → 0.1.0

package/README.md

@@ -9,7 +9,7 @@
 implemented using TypeGPU. It composes GPU kernels into the layers a model is
 made of, attention, convolutions, norms, matmuls and builds common
 architectures out of them. The kernels are plain TypeGPU code compiled to
-WGSL at build time, so there is no WASM binary to download at runtime, and
+WGSL at runtime, so there is no WASM binary to download, and
 the same code runs anywhere WebGPU does: the browser, React Native and
 Electron.
 
@@ -34,7 +34,7 @@ It ships as one package, `runntime`, with two entry points:
 - **One function per task** - `createObjectDetector()` hands you a loaded
   model. Call `detectObjects(image)`, get boxes back. The `zoo` api is meant to be straight-forward and just work.
 - **No WASM to download** - the kernels are TypeGPU code compiled
-  to WGSL at build time, so there is no WASM binary shipped alongside your
+  to WGSL at runtime, so there is no WASM binary shipped alongside your
   app.
 - **Same code everywhere WebGPU runs** - the browser, React Native (through
   [react-native-webgpu](https://github.com/wcandillon/react-native-webgpu))

package/dist/{classifier-C9dvSwSY.mjs → classifier-ByH6OI6h.mjs}

@@ -41864,11 +41864,14 @@ const MOONSHINE_STREAMING_TINY = precisions("moonshine-streaming", "tiny", {
 });
 const PRIVACY_FILTER_INT8 = { modelPath: weights("privacy-filter", "base", "int8g640") };
 const YOLO26_N = { modelPath: weights("yolo26", "n", "f16") };
+const YOLO26_S = { modelPath: weights("yolo26", "s", "f16") };
 const YOLO26_M = { modelPath: weights("yolo26", "m", "f16") };
 const MOBILENETV4_SMALL = { modelPath: weights("mobilenetv4", "small", "f16") };
 const YOLO26_SEG_N = { modelPath: weights("yolo26-seg", "n", "f16") };
+const YOLO26_SEG_S = { modelPath: weights("yolo26-seg", "s", "f16") };
 const YOLO26_SEG_M = { modelPath: weights("yolo26-seg", "m", "f16") };
 const YOLO26_POSE_N = { modelPath: weights("yolo26-pose", "n", "f16") };
+const YOLO26_POSE_S = { modelPath: weights("yolo26-pose", "s", "f16") };
 const YOLO26_POSE_M = { modelPath: weights("yolo26-pose", "m", "f16") };
 const DEPTHART_S = { modelPath: weights("depthart", "s", "f16") };
 const DEPTHART_B = { modelPath: weights("depthart", "b", "f16") };
@@ -41899,16 +41902,20 @@ YOLO26: {
 		DEFAULT: YOLO26_N,
 		/** The nano size, 5 MB. */
 		N: YOLO26_N,
-		/** The medium size, 41 MB. More accurate, slower. */
+		/** The small size, 18 MB. */
+		S: YOLO26_S,
+		/** The medium size, 39 MB. More accurate, slower. */
 		M: YOLO26_M
 	} },
 	instanceSegmentation: { 
 	/** YOLO26 segment, COCO-80 classes with a mask per object, f16 weights. */
 YOLO26_SEG: {
 		DEFAULT: YOLO26_SEG_N,
-		/** The nano size, 6 MB. */
+		/** The nano size, 5 MB. */
 		N: YOLO26_SEG_N,
-		/** The medium size, 48 MB. More accurate, slower. */
+		/** The small size, 20 MB. */
+		S: YOLO26_SEG_S,
+		/** The medium size, 45 MB. More accurate, slower. */
 		M: YOLO26_SEG_M
 	} },
 	keypointDetection: { 
@@ -41917,7 +41924,9 @@ YOLO26_POSE: {
 		DEFAULT: YOLO26_POSE_N,
 		/** The nano size, 6 MB. */
 		N: YOLO26_POSE_N,
-		/** The medium size, 44 MB. More accurate, slower. */
+		/** The small size, 20 MB. */
+		S: YOLO26_POSE_S,
+		/** The medium size, 42 MB. More accurate, slower. */
 		M: YOLO26_POSE_M
 	} },
 	depthEstimation: { 

package/dist/zoo/index.d.mts

@@ -175,7 +175,11 @@ export declare const models: {
       readonly N: {
         modelPath: string;
       };
-      /** The medium size, 41 MB. More accurate, slower. */
+      /** The small size, 18 MB. */
+      readonly S: {
+        modelPath: string;
+      };
+      /** The medium size, 39 MB. More accurate, slower. */
       readonly M: {
         modelPath: string;
       };
@@ -187,11 +191,15 @@ export declare const models: {
       readonly DEFAULT: {
         modelPath: string;
       };
-      /** The nano size, 6 MB. */
+      /** The nano size, 5 MB. */
       readonly N: {
         modelPath: string;
       };
-      /** The medium size, 48 MB. More accurate, slower. */
+      /** The small size, 20 MB. */
+      readonly S: {
+        modelPath: string;
+      };
+      /** The medium size, 45 MB. More accurate, slower. */
       readonly M: {
         modelPath: string;
       };
@@ -207,7 +215,11 @@ export declare const models: {
       readonly N: {
         modelPath: string;
       };
-      /** The medium size, 44 MB. More accurate, slower. */
+      /** The small size, 20 MB. */
+      readonly S: {
+        modelPath: string;
+      };
+      /** The medium size, 42 MB. More accurate, slower. */
       readonly M: {
         modelPath: string;
       };

package/dist/zoo/index.mjs

@@ -1,4 +1,4 @@
-import { $ as initRunntime, A as createMinilmEmbedCore, B as Linear, C as decodeTokens, F as createOpfsCache, H as Module, J as derivedTensor, K as warmUp, L as Conv1d, M as EMBEDDING_DIM, N as MINILM_L6, O as MinilmModel, P as models, Q as gpuExecutor, R as Embedding, T as createDetector, U as ModuleList, V as MultiHeadAttention, W as Parameter, Y as fromSafetensors, Z as evalValues, _ as buildRopeTables, a as MOBILENETV4_CROP_FRACTION, at as writeRows, c as createDetector$1, ct as add, d as move, dt as mul, f as moveTiedEmbedding, ft as silu, g as MoonshineDecoder, h as transformDecoder, ht as supportsF16, it as transpose, l as createTranscriber$1, lt as asinh, m as take, mt as defaultRoot, n as IMAGENET_MEAN, nt as gatherRowsFrom, o as DEPTHART_CONFIGS, ot as argmaxDot, p as qualify, pt as tensor, q as bufferSource, r as IMAGENET_STD, rt as slice, s as createEstimator, st as matmul, t as createClassifier, tt as chunk, u as fuseQkv, ut as gelu, v as splitIntoChunks, w as moonshineTokenizerAsset, z as LayerNorm } from "../classifier-C9dvSwSY.mjs";
+import { $ as initRunntime, A as createMinilmEmbedCore, B as Linear, C as decodeTokens, F as createOpfsCache, H as Module, J as derivedTensor, K as warmUp, L as Conv1d, M as EMBEDDING_DIM, N as MINILM_L6, O as MinilmModel, P as models, Q as gpuExecutor, R as Embedding, T as createDetector, U as ModuleList, V as MultiHeadAttention, W as Parameter, Y as fromSafetensors, Z as evalValues, _ as buildRopeTables, a as MOBILENETV4_CROP_FRACTION, at as writeRows, c as createDetector$1, ct as add, d as move, dt as mul, f as moveTiedEmbedding, ft as silu, g as MoonshineDecoder, h as transformDecoder, ht as supportsF16, it as transpose, l as createTranscriber$1, lt as asinh, m as take, mt as defaultRoot, n as IMAGENET_MEAN, nt as gatherRowsFrom, o as DEPTHART_CONFIGS, ot as argmaxDot, p as qualify, pt as tensor, q as bufferSource, r as IMAGENET_STD, rt as slice, s as createEstimator, st as matmul, t as createClassifier, tt as chunk, u as fuseQkv, ut as gelu, v as splitIntoChunks, w as moonshineTokenizerAsset, z as LayerNorm } from "../classifier-ByH6OI6h.mjs";
 //#region src/core/lifetime.ts
 function createResourceScope() {
 	const tracked = [];

package/dist/zoo/transformers/index.mjs

@@ -1,4 +1,4 @@
-import { $ as initRunntime, D as V2_NER_CLASS_NAMES, E as loadPrivacyFilterModel, F as createOpfsCache, G as inGpuErrorScopes, I as createReplayCache, N as MINILM_L6, O as MinilmModel, P as models, S as configFromCheckpoint, X as httpRangeSource, Y as fromSafetensors, b as MOONSHINE_TINY, c as createDetector, et as toArray, ht as supportsF16, i as MOBILENETV4_CONV_S, j as fillBucketInputs, k as MINILM_BUCKETS, l as createTranscriber, mt as defaultRoot, q as bufferSource, s as createEstimator, t as createClassifier, x as assertConfigMatches, y as MOONSHINE_BASE } from "../../classifier-C9dvSwSY.mjs";
+import { $ as initRunntime, D as V2_NER_CLASS_NAMES, E as loadPrivacyFilterModel, F as createOpfsCache, G as inGpuErrorScopes, I as createReplayCache, N as MINILM_L6, O as MinilmModel, P as models, S as configFromCheckpoint, X as httpRangeSource, Y as fromSafetensors, b as MOONSHINE_TINY, c as createDetector, et as toArray, ht as supportsF16, i as MOBILENETV4_CONV_S, j as fillBucketInputs, k as MINILM_BUCKETS, l as createTranscriber, mt as defaultRoot, q as bufferSource, s as createEstimator, t as createClassifier, x as assertConfigMatches, y as MOONSHINE_BASE } from "../../classifier-ByH6OI6h.mjs";
 import tgpu from "typegpu";
 import { AutoConfig, PreTrainedModel, Tensor } from "@huggingface/transformers";
 //#region src/zoo/transformers/minilm/model.ts
@@ -434,7 +434,8 @@ function selectTopAnchors(levels, opts) {
 	}
 	const order = topIndices(best, maxDet);
 	const count = Math.min(maxDet, total);
-	const logits = new Float32Array(count * numClasses);
+	const outClasses = numClasses + 1;
+	const logits = new Float32Array(count * outClasses);
 	const boxes = new Float32Array(count * 4);
 	for (let k = 0; k < count; k++) {
 		let idx = order[k];
@@ -445,7 +446,7 @@ function selectTopAnchors(levels, opts) {
 		}
 		const { data, h, w, stride } = levels[level];
 		const hw = h * w;
-		for (let c = 0; c < numClasses; c++) logits[k * numClasses + c] = data[(4 + c) * hw + idx];
+		for (let c = 0; c < numClasses; c++) logits[k * outClasses + c] = data[(4 + c) * hw + idx];
 		const ax = idx % w + .5;
 		const ay = Math.floor(idx / w) + .5;
 		const l = data[idx];
@@ -461,7 +462,8 @@ function selectTopAnchors(levels, opts) {
 	return {
 		logits,
 		boxes,
-		count
+		count,
+		outClasses
 	};
 }
 //#endregion
@@ -494,7 +496,7 @@ var RunntimeYolosForObjectDetection = class extends PreTrainedModel {
 			logits: new Tensor("float32", top.logits, [
 				1,
 				top.count,
-				numClasses
+				top.outClasses
 			]),
 			pred_boxes: new Tensor("float32", top.boxes, [
 				1,
@@ -515,10 +517,15 @@ function yolo26Variant(modelId) {
 /** The hosted f16 exports, by variant. */
 const YOLO26_WEIGHTS_URLS = {
 	n: models.objectDetection.YOLO26.N.modelPath,
+	s: models.objectDetection.YOLO26.S.modelPath,
 	m: models.objectDetection.YOLO26.M.modelPath
 };
 /** Every hub id with a hosted export. */
-const YOLO26_MODEL_IDS = ["onnx-community/yolo26n-ONNX", "onnx-community/yolo26m-ONNX"];
+const YOLO26_MODEL_IDS = [
+	"onnx-community/yolo26n-ONNX",
+	"onnx-community/yolo26s-ONNX",
+	"onnx-community/yolo26m-ONNX"
+];
 /** `weightsUrl` defaults to the hosted export of the variant in the hub
 *  id; other variants need one. */
 function yolo26Loader(weightsUrl, opts = {}) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runntime",
-  "version": "0.0.1-alpha.1",
+  "version": "0.1.0",
   "description": "AI models on the web, built with TypeGPU.",
   "license": "MIT",
   "repository": {
@@ -45,7 +45,7 @@
   "scripts": {
     "build": "tsdown",
     "typecheck": "tsc --noEmit",
-    "prepack": "pnpm run build"
+    "prepack": "node ../../scripts/sync-readme.mjs && pnpm run build"
   },
   "dependencies": {
     "js-tiktoken": "^1.0.20"