runntime 0.1.0 → 0.2.0

package/dist/zoo/transformers/index.d.mts

@@ -1,4 +1,4 @@
-import { t as Yolo26Variant } from "../../config-UyxIidZc.mjs";
+import { t as Yolo26Variant } from "../../config-DZyZ2TdH.mjs";
 import { TgpuRoot } from "typegpu";
 import { PreTrainedModel, PretrainedConfig } from "@huggingface/transformers";
 //#region src/zoo/transformers/registry.d.ts
@@ -13,13 +13,13 @@ export declare function registerRunntimeBackend(opts?: RegisterRunntimeBackendOp
 export declare function unregisterRunntimeBackend(): void;
 //#endregion
 //#region src/zoo/transformers/init.d.ts
-/** One-call plugin setup: creates the WebGPU device, points runntime/core at it
- *  and patches transformers.js. Requests subgroups, which the fast matmul
- *  and attention kernels need, and shader-f16, which yolo26 and depthart
- *  require and the other models use when present. Both are optional: a
- *  device lacking either still loads the f32-capable models. Options go
- *  straight to registerRunntimeBackend(), so `models` and `fallbackToOnnx` work
- *  here too. Returns the root for apps that also use TypeGPU directly. */
+/** One-call plugin setup: creates the WebGPU device, runs the models on it
+ *  and patches transformers.js. Requests `subgroups`, which makes matmul and
+ *  attention faster, and `shader-f16`, which the YOLO26 and DepthART models
+ *  need and the others use when present. A device without them still loads
+ *  every model that runs in f32. Options go straight to
+ *  registerRunntimeBackend(), so `models` and `fallbackToOnnx` work here too.
+ *  Returns the root for apps that also use TypeGPU directly. */
 export declare function initRunntimeBackend(opts?: RegisterRunntimeBackendOpts): Promise<TgpuRoot>;
 //#endregion
 //#region src/zoo/transformers/runntimeModel.d.ts
@@ -51,8 +51,8 @@ export declare function yolo26Loader(weightsUrl?: string, opts?: {
 //#endregion
 //#region src/zoo/transformers/depthart/loader.d.ts
 /** `weightsUrl`: a safetensors export of a Fengxue93/DepthART relative
- *  checkpoint (tools/export_weights_depthart.py); defaults to the hosted
- *  export for the hub id. The variant is read from the file. */
+ *  checkpoint. Defaults to the hosted export for the hub id. The variant is
+ *  read from the file. */
 export declare function depthartLoader(weightsUrl?: string, opts?: {
   onProgress?: (name: string, doneBytes: number, totalBytes: number) => void;
 }): RunntimeModelLoader;

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-ByH6OI6h.mjs";
+import { At as defaultRoot, D as MINILM_BUCKETS, E as MinilmModel, L as models, Nt as RunntimeError, Pt as isRunntimeError, Q as bufferSource, R as createOpfsCache, S as configFromCheckpoint, T as V2_NER_CLASS_NAMES, X as inGpuErrorScopes, at as toArray, b as MOONSHINE_TINY, c as createDetector, et as fromSafetensors, i as MOBILENETV4_CONV_S, it as initRunntime, j as MINILM_L6, jt as supportsF16, k as fillBucketInputs, l as createTranscriber, s as createEstimator, t as createClassifier, tt as httpRangeSource, w as loadPrivacyFilterModel, x as assertConfigMatches, y as MOONSHINE_BASE, z as createReplayCache } from "../../classifier-C6ZR3cyJ.mjs";
 import tgpu from "typegpu";
 import { AutoConfig, PreTrainedModel, Tensor } from "@huggingface/transformers";
 //#region src/zoo/transformers/minilm/model.ts
@@ -138,7 +138,7 @@ async function loadCachedWeights(url, key) {
 //#region src/zoo/transformers/minilm/loader.ts
 /** The hosted f16 export; loadStateDict fuses qkv and folds the token-type
 *  embedding while loading it. */
-const MINILM_WEIGHTS_URL = models.textEmbedding.ALL_MINILM_L6_V2.DEFAULT.modelPath;
+const MINILM_WEIGHTS_URL = models.embedding.ALL_MINILM_L6_V2.DEFAULT.modelPath;
 const MINILM_MODEL_IDS = ["Xenova/all-MiniLM-L6-v2", "sentence-transformers/all-MiniLM-L6-v2"];
 function minilmLoader(weightsUrl = MINILM_WEIGHTS_URL) {
 	return { async load(_modelId, config) {
@@ -296,13 +296,11 @@ function moonshinePreset(config) {
 		cfg
 	};
 }
+/** The hub tokenizer decodes the text, so the transcriber's text stays empty. */
 function idsOnlyTokenizer(cfg) {
 	return {
-		version: 0,
-		bosId: cfg.bos,
 		eosId: cfg.eos,
-		specialIds: [cfg.bos, cfg.eos],
-		vocab: new Array(cfg.vocab).fill("")
+		decode: () => ""
 	};
 }
 function moonshineLoader(weightsUrl, opts = {}) {
@@ -589,8 +587,8 @@ const DEPTHART_WEIGHTS_URLS = {
 };
 const DEPTHART_MODEL_IDS = Object.keys(DEPTHART_WEIGHTS_URLS);
 /** `weightsUrl`: a safetensors export of a Fengxue93/DepthART relative
-*  checkpoint (tools/export_weights_depthart.py); defaults to the hosted
-*  export for the hub id. The variant is read from the file. */
+*  checkpoint. Defaults to the hosted export for the hub id. The variant is
+*  read from the file. */
 function depthartLoader(weightsUrl, opts = {}) {
 	return { async load(modelId, config) {
 		const root = defaultRoot();
@@ -630,7 +628,7 @@ function registerRunntimeBackend(opts = {}) {
 			const config = await AutoConfig.from_pretrained(modelId, options);
 			return await loader.load(modelId, config);
 		} catch (e) {
-			if (!fallbackToOnnx) throw new Error(`runntime backend: loading '${modelId}' failed. Pass fallbackToOnnx: true to load it on ONNX instead.`, { cause: e });
+			if (!fallbackToOnnx) throw new RunntimeError(isRunntimeError(e) ? e.code : "LOAD_FAILED", `runntime backend: loading '${modelId}' failed. Pass fallbackToOnnx: true to load it on ONNX instead.`, { cause: e });
 			console.warn(`runntime backend: loading '${modelId}' failed, falling back to ONNX`, e);
 		}
 		else if (!fallbackToOnnx) {
@@ -650,13 +648,13 @@ function unregisterRunntimeBackend() {
 }
 //#endregion
 //#region src/zoo/transformers/init.ts
-/** One-call plugin setup: creates the WebGPU device, points runntime/core at it
-*  and patches transformers.js. Requests subgroups, which the fast matmul
-*  and attention kernels need, and shader-f16, which yolo26 and depthart
-*  require and the other models use when present. Both are optional: a
-*  device lacking either still loads the f32-capable models. Options go
-*  straight to registerRunntimeBackend(), so `models` and `fallbackToOnnx` work
-*  here too. Returns the root for apps that also use TypeGPU directly. */
+/** One-call plugin setup: creates the WebGPU device, runs the models on it
+*  and patches transformers.js. Requests `subgroups`, which makes matmul and
+*  attention faster, and `shader-f16`, which the YOLO26 and DepthART models
+*  need and the others use when present. A device without them still loads
+*  every model that runs in f32. Options go straight to
+*  registerRunntimeBackend(), so `models` and `fallbackToOnnx` work here too.
+*  Returns the root for apps that also use TypeGPU directly. */
 async function initRunntimeBackend(opts = {}) {
 	const root = await tgpu.init({ device: { optionalFeatures: ["subgroups", "shader-f16"] } });
 	initRunntime(root);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runntime",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "AI models on the web, built with TypeGPU.",
   "license": "MIT",
   "repository": {
@@ -45,14 +45,19 @@
   "scripts": {
     "build": "tsdown",
     "typecheck": "tsc --noEmit",
-    "prepack": "node ../../scripts/sync-readme.mjs && pnpm run build"
+    "prepack": "node ../../scripts/sync-readme.mjs && pnpm run build",
+    "test": "vitest run",
+    "test:unit": "vitest run --project unit",
+    "test:gpu": "vitest run --project gpu",
+    "bench:ab": "node bench/cli.ts"
   },
   "dependencies": {
+    "@huggingface/tokenizers": "^0.1.3",
     "js-tiktoken": "^1.0.20"
   },
   "peerDependencies": {
-    "typegpu": "^0.12.5",
-    "@huggingface/transformers": "^4.2.0"
+    "@huggingface/transformers": "^4.2.0",
+    "typegpu": "^0.12.5"
   },
   "peerDependenciesMeta": {
     "@huggingface/transformers": {
@@ -62,9 +67,12 @@
   "devDependencies": {
     "@huggingface/transformers": "^4.2.0",
     "@types/node": "^25.9.2",
+    "@vitest/browser-playwright": "^5.0.4",
     "@webgpu/types": "^0.1.60",
+    "playwright": "^1.64.0",
     "tsdown": "^0.23.0",
+    "typegpu": "^0.12.5",
     "unplugin-typegpu": "^0.12.3",
-    "typegpu": "^0.12.5"
+    "vitest": "^5.0.4"
   }
 }