@simulatte/doppler 0.1.8 → 0.1.9

package/CHANGELOG.md

@@ -6,22 +6,35 @@ This changelog is package-facing and release-oriented. Entries before `0.1.7`
 were retrofitted from package version history, release commits, and release
 docs so the `0.1.x` line has one conventional npm-visible history surface.
 
-## [0.1.8] - 2026-03-13
+## [0.1.8] - 2026-03-18
 
 ### Changed
 
 - Simplified demo to show only verified Q4K models (Gemma 3 270M, Gemma 3 1B).
   Hidden Translate, Diffusion, and Embedding tabs until models are ready.
+- Split demo monolith (6,680 lines) into focused modules: core, generation,
+  storage, translate, diagnostics, routing, utils.
 - Trimmed hosted HF registry and quickstart registry to the two verified models.
 - Aligned catalog, HF registry, and quickstart registry to the canonical
   external support registry as single source of truth for HF revisions.
+- Renamed all `.mjs` tool scripts to `.js` to match `"type": "module"` convention.
+- Switched WebGPU optional dependency from `@simulatte/webgpu` to `webgpu ^0.3.8`.
+- Pruned unused `verify:*` npm scripts for models no longer in the active set.
+- Updated release-claim policy with newly verified models (LFM2, Qwen 3.5,
+  TranslateGemma variants).
 
 ### Fixed
 
 - Fixed Qwen 3.5 conversion configs using wrong model preset (`qwen3` instead
   of `qwen3_5`), which caused support matrix check failures.
+- Fixed Qwen mRoPE conflation: `ropeInterleaved` was incorrectly set from
+  `mropeInterleaved`, forcing adjacent-pair RoPE rotation on Qwen models.
 - Fixed catalog lifecycle metadata inconsistencies: corrected `local`, `hf`,
   `curated`, and `demo` fields to match actual artifact availability.
+- Fixed GPU-dependent unit tests failing in non-GPU environments by adding
+  proper GPU readiness probes with clear skip reasons.
+- Fixed kernel-ref digest registry drift (222 vs 224 entries).
+- Fixed stale vendor benchmark fixture hashes after compare-engines config update.
 - Removed failing and unverified models from demo visibility (TranslateGemma 4B,
   EmbeddingGemma 300M with broken HF manifest, Qwen 3.5 0.8B/2B, F16 variant).
 

package/README.md

@@ -2,6 +2,8 @@
 
 Inference and training on raw WebGPU. Pure JS + WGSL.
 
+**[Try the live demo](https://d4da.com)** | **[npm](https://www.npmjs.com/package/@simulatte/doppler)** | **[docs](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)**
+
 ![Phase-latency comparison on one workload across models](https://raw.githubusercontent.com/clocksmith/doppler/main/benchmarks/vendors/results/compare_1b_multi-workload_favorable_phases.svg)
 
 ## Quick start
@@ -28,8 +30,6 @@ Registry IDs resolve to hosted RDRR artifacts from `Clocksmith/rdrr` by default.
 npm install @simulatte/doppler
 ```
 
-**[Live Demo](https://d4da.com)** · **[npm](https://www.npmjs.com/package/@simulatte/doppler)** · **[docs](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)** · **[Project site](https://simulatte.world)**
-
 ## Why Doppler
 
 **JS → WGSL → WebGPU.** Direct JavaScript orchestration into native WebGPU kernels, avoiding ONNX runtimes, WASM blobs, and bridge layers.
@@ -46,6 +46,28 @@ Snapshot artifacts:
 - [g3-1b-p064-d064-t0-k1.compare.json](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/fixtures/g3-1b-p064-d064-t0-k1.compare.json)
 - [lfm2-5-1-2b-p064-d064-t0-k1.compare.json](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/fixtures/lfm2-5-1-2b-p064-d064-t0-k1.compare.json)
 
+## Supported models
+
+All models below are verified with deterministic greedy decoding on WebGPU hardware.
+Registry IDs resolve to hosted RDRR artifacts automatically.
+
+| Model | Registry ID | Quant | Params |
+| --- | --- | --- | --- |
+| Gemma 3 270M IT | `gemma3-270m` | Q4K | 270M |
+| Gemma 3 1B IT | `gemma3-1b` | Q4K | 1B |
+| Gemma 3 1B IT (F16) | `gemma-3-1b-it-f16-af32` | F16 | 1B |
+| TranslateGemma 4B IT | `translategemma-4b-it-q4k-ehf16-af32` | Q4K | 4B |
+| TranslateGemma 4B 1B EN-ES | `translategemma-4b-1b-enes-q4k-ehf16-af32` | Q4K | 1B |
+| EmbeddingGemma 300M | `google-embeddinggemma-300m-q4k-ehf16-af32` | Q4K | 300M |
+| Qwen 3.5 0.8B | `qwen-3-5-0-8b-q4k-ehaf16` | Q4K | 0.8B |
+| Qwen 3.5 2B | `qwen-3-5-2b-q4k-ehaf16` | Q4K | 2B |
+| LFM2.5 1.2B Instruct | `lfm2-5-1-2b-instruct-q4k-ehf16-af32` | Q4K | 1.2B |
+
+Additional model families (Llama 3, DeepSeek, Gemma 4 MoE, Mixtral, and others) have conversion
+configs ready but are not yet cataloged. See the full
+[model support matrix](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md)
+for details.
+
 ## Under the hood
 
 - Sharded weight loading via OPFS moves multi-GB weights into VRAM without blocking the main thread.
@@ -85,10 +107,7 @@ for await (const token of doppler('Hello', { model: 'gemma3-270m' })) {
 - First-run workflow: [docs/getting-started.md](https://github.com/clocksmith/doppler/blob/main/docs/getting-started.md)
 - Runtime config contract: [docs/config.md](https://github.com/clocksmith/doppler/blob/main/docs/config.md)
 - Architecture: [docs/architecture.md](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md)
-- Generated model support table: [docs/model-support-matrix.md](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md)
-
-Current model support is generated from the catalog and conversion registry.
-See [docs/model-support-matrix.md](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md) for the canonical verified, failing, and unverified status table.
+- Model support matrix: [docs/model-support-matrix.md](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md)
 
 ## Environment requirements
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@simulatte/doppler",
-  "version": "0.1.8",
+  "version": "0.1.9",
   "description": "Browser-native WebGPU inference engine for local intent and inference loops",
   "main": "src/index.js",
   "types": "src/index.d.ts",
@@ -75,6 +75,7 @@
     "external:rdrr:index": "node tools/sync-external-rdrr-index.js",
     "external:rdrr:index:check": "node tools/sync-external-rdrr-index.js --check",
     "external:support:sync": "node tools/sync-external-support-registry.js",
+    "external:support:promote": "node tools/sync-external-support-registry.js --source-support-file models/catalog.json",
     "external:support:check": "node tools/sync-external-support-registry.js --check",
     "catalog:sync:external": "node tools/sync-catalog-from-external-support.js",
     "catalog:sync:external:check": "node tools/sync-catalog-from-external-support.js --check",
@@ -155,12 +156,13 @@
     "tools/convert-safetensors-node.js"
   ],
   "devDependencies": {
-    "@huggingface/transformers": "^3.8.1",
+    "@huggingface/transformers": "4.0.0-next.8",
     "jest": "^30.2.0",
     "onnxruntime-web": "^1.24.1",
     "playwright": "^1.58.2"
   },
   "optionalDependencies": {
-    "@simulatte/webgpu": "0.1.x"
+    "@simulatte/webgpu": "0.x.x",
+    "webgpu": "^0.3.8"
   }
 }

package/src/client/doppler-api.browser.js

@@ -165,6 +165,12 @@ function createModelHandle(pipeline, resolved) {
       prefillKV(prompt, options = {}) {
         return pipeline.prefillKVOnly(prompt, options);
       },
+      prefillWithLogits(prompt, options = {}) {
+        return pipeline.prefillWithLogits(prompt, options);
+      },
+      decodeStepLogits(currentIds, options = {}) {
+        return pipeline.decodeStepLogits(currentIds, options);
+      },
       generateWithPrefixKV(prefix, prompt, options = {}) {
         return pipeline.generateWithPrefixKV(prefix, prompt, options);
       },

package/src/client/doppler-api.d.ts

@@ -2,6 +2,7 @@ import type { RDRRManifest } from '../formats/rdrr/index.js';
 import type { GenerateOptions, KVCacheSnapshot } from '../generation/index.js';
 import type { ChatMessage } from '../inference/pipelines/text/chat-format.js';
 import type { LoRAManifest } from '../adapters/lora-loader.js';
+import type { LogitsStepResult, PrefillResult } from '../inference/pipelines/text/types.d.ts';
 
 export interface DopplerLoadProgress {
   phase: 'resolve' | 'manifest' | 'load' | 'ready';
@@ -43,6 +44,8 @@ export interface DopplerModel {
   readonly deviceInfo: Record<string, unknown> | null;
   readonly advanced: {
     prefillKV(prompt: string, options?: GenerateOptions): Promise<KVCacheSnapshot>;
+    prefillWithLogits(prompt: string | ChatMessage[] | { messages: ChatMessage[] }, options?: GenerateOptions): Promise<PrefillResult>;
+    decodeStepLogits(currentIds: number[], options?: GenerateOptions): Promise<LogitsStepResult>;
     generateWithPrefixKV(
       prefix: KVCacheSnapshot,
       prompt: string,

package/src/client/doppler-api.js

@@ -199,6 +199,12 @@ function createModelHandle(pipeline, resolved) {
       prefillKV(prompt, options = {}) {
         return pipeline.prefillKVOnly(prompt, options);
       },
+      prefillWithLogits(prompt, options = {}) {
+        return pipeline.prefillWithLogits(prompt, options);
+      },
+      decodeStepLogits(currentIds, options = {}) {
+        return pipeline.decodeStepLogits(currentIds, options);
+      },
       generateWithPrefixKV(prefix, prompt, options = {}) {
         return pipeline.generateWithPrefixKV(prefix, prompt, options);
       },
@@ -272,6 +278,9 @@ export function doppler(prompt, options) {
 
 doppler.load = load;
 doppler.text = async function text(prompt, options) {
+  if (!options || typeof options !== 'object' || options.model == null) {
+    throw new Error('doppler.text() requires options.model.');
+  }
   assertNoLoadAffectingOptions('doppler.text()', options);
   return collectText(doppler(prompt, options));
 };
@@ -299,14 +308,14 @@ doppler.evict = async function evict(model) {
   if (!cached) {
     return false;
   }
-  convenienceModelCache.delete(resolved.modelId);
   await cached.unload();
+  convenienceModelCache.delete(resolved.modelId);
   return true;
 };
 doppler.evictAll = async function evictAll() {
   const cached = [...convenienceModelCache.values()];
   convenienceModelCache.clear();
-  await Promise.all(cached.map((entry) => entry.unload()));
+  await Promise.allSettled(cached.map((entry) => entry.unload()));
 };
 doppler.listModels = async function listModels() {
   const models = await listQuickstartModels();

package/src/client/doppler-registry.js

@@ -1,4 +1,5 @@
 import { getCdnBasePath } from '../storage/download-types.js';
+import { buildHfResolveBaseUrl } from '../utils/hf-resolve-url.js';
 import { loadJson } from '../utils/load-json.js';
 
 let registryPromise = null;
@@ -80,9 +81,6 @@ export function buildQuickstartModelBaseUrl(entry, options = {}) {
   }
   const cdnBasePath = typeof options.cdnBasePath === 'string' && options.cdnBasePath.length > 0
     ? options.cdnBasePath
-    : (getCdnBasePath() || 'https://huggingface.co');
-  const revision = entry.hf.revision || 'main';
-  const base = cdnBasePath.replace(/\/$/, '');
-  const path = entry.hf.path.replace(/^\/+/, '');
-  return `${base}/${entry.hf.repoId}/resolve/${revision}/${path}`;
+    : getCdnBasePath();
+  return buildHfResolveBaseUrl(entry.hf, { cdnBasePath });
 }

package/src/client/doppler-registry.json

@@ -19,6 +19,22 @@
         "revision": "ca6f0dbdf3882d3893a65cf48f2bb6f1520df162",
         "path": "models/gemma-3-270m-it-q4k-ehf16-af32"
       }
+    },
+    {
+      "modelId": "google-embeddinggemma-300m-q4k-ehf16-af32",
+      "aliases": [
+        "embeddinggemma-300m",
+        "google/embeddinggemma-300m",
+        "google-embeddinggemma-300m-wq4k-ef16"
+      ],
+      "modes": [
+        "embedding"
+      ],
+      "hf": {
+        "repoId": "Clocksmith/rdrr",
+        "revision": "7e79c466d54455bd370c81685956ea9abae0fd30",
+        "path": "models/google-embeddinggemma-300m-q4k-ehf16-af32"
+      }
     }
   ]
 }

package/src/config/kernels/kernel-ref-digests.js

@@ -59,8 +59,8 @@ export const KERNEL_REF_CONTENT_DIGESTS = Object.freeze({
   "cross_entropy_loss.wgsl#main": "5a48087bdec94184432c90ce5b345e1eadbdfcb13b9793ecee8052bc7392239c",
   "depthwise_conv2d_f16.wgsl#main": "f7f093a7e6623ed17a675bac729149e94718aece916416966eaf03c1d6939f2a",
   "depthwise_conv2d.wgsl#main": "cf14cb40d282ad4d4fab160109b97eaeaf12aab62579b73324ac485ac75155b0",
-  "dequant_f16_out_vec4.wgsl#main_vec4": "61c20e6c71c1c8421b4ec202dbd26292a6300587bd44c314f2a6c6d9d9442c3a",
-  "dequant_f16_out.wgsl#main": "94d61843d56f9a3bbc6b7c2b95dc6ecbba3f6a262b2c4086a076f69a8c38ccae",
+  "dequant_f16_out_vec4.wgsl#main_vec4": "ff729cc220ba5425e17c4c537a9993f25b6541046b6c2553d2a43a8b40ed2ce9",
+  "dequant_f16_out.wgsl#main": "caed21e420cbace78d3203548962a5ec3fc36980f153ae775f6a91a31af97d3a",
   "dequant_f16_rowwise.wgsl#main": "f5bf7cef950b52d65cee6121dbaa176244d3221045b3b6386b3be47f23ce17dc",
   "dequant_f32_rowwise.wgsl#main": "e73606e1b47e1191203a210bececa8a597bcab8bcc535146718afa6a021cab0d",
   "dequant_mxfp4_expert_f16.wgsl#main_expert": "96af52551ac40e1b86121a528a3ffaba835c5d0419e06407fed80353d46b17e1",
@@ -69,10 +69,10 @@ export const KERNEL_REF_CONTENT_DIGESTS = Object.freeze({
   "dequant_mxfp4.wgsl#main": "885a5f752b684c6ca0bb10e3a1846a396eef14d2158e8c8ad31bd1dd4c74b9ef",
   "dequant_q6k.wgsl#main": "be0aed027932d8b7dd1e92d0090ced39e4df8be724acf290f52db0004be9a35e",
   "dequant_q8_0.wgsl#main": "ff5f800da963b0502a9ffab723cbcac0bbb5eb9a02898afc2aba2db215a58da7",
-  "dequant_shared_vec4.wgsl#main_vec4": "aaf330e814cbd4c2838f07639025e414542ec2d0ddbb34913d7f10d28253baf2",
-  "dequant_shared.wgsl#main": "d83d83a5d2bbcd29a06f67fb41496edaa45216438ff0dd177e7ec33af08104a4",
-  "dequant_subgroup.wgsl#main": "6b6296d8b060a9b16e7b3c2bee99dbbcbbf9cc3d557f1e2b000f7f9bfc020309",
-  "dequant_subgroup.wgsl#main_vec4": "d9de6ea645f558bcdf6db622fe146710e143fc4c2bb56f130342d62858eb456c",
+  "dequant_shared_vec4.wgsl#main_vec4": "24820dae36f6669a33f22b428df03791d9c700944c5ae33bd8c88e8cbeffd103",
+  "dequant_shared.wgsl#main": "e21284b5b70d4ac88d7c151760e451c2006705f1ea617b3db7f89994af4cc7df",
+  "dequant_subgroup.wgsl#main": "cbc2d86a5a2234b4c1691d5df02279263be7a66a1d4a2ad4aec1845a26baa9c9",
+  "dequant_subgroup.wgsl#main_vec4": "9e044bd0f44e73872dd8d8aa467e802c5471de86a2044de2cf8efc726e5a1182",
   "energy_eval_f16.wgsl#main": "09223ae193593f3555866a3acfe76ca35442ef4f3967cae376bdcc211f3054b3",
   "energy_eval.wgsl#main": "e10d9572397ebece5275aecd907cba5970f6a5c3744dd8b982677efb8982bdd2",
   "energy_quintel_grad_f16.wgsl#main": "eb87ed8592b46b0a4d866c245b664cadb2bca016f72419e763402a6a721c4951",
@@ -142,9 +142,9 @@ export const KERNEL_REF_CONTENT_DIGESTS = Object.freeze({
   "matmul_gemv_subgroup_f16a.wgsl#main_multicol": "c8e86ecbbefa27a3b7366af676d89a992c2e951329cdf19abb57b9c90144379e",
   "matmul_gemv_subgroup_f16a.wgsl#main_vec4": "f227a403cdf9717dd68224c9ea55708ffe14c618d8146f5d48b42af0f253df29",
   "matmul_gemv_subgroup_f16a.wgsl#main_vec4_cols8": "9e7aba97a6cf199b3f574166e295ea051ebd59e308b5f6f2ce5a4de2d04963ce",
-  "matmul_gemv_subgroup.wgsl#main": "ff526a7e70d6933f5c7a59aee476cc9c5cf123903222ed2e74f555e86e1ac757",
-  "matmul_gemv_subgroup.wgsl#main_multicol": "96c38c15e6fed0d7efdc5cd094db5843a8e8ddfe01eee3bc7322fa555dacf3d0",
-  "matmul_gemv_subgroup.wgsl#main_vec4": "3cee3bed453b40c5564a751d2a917649e10ad52f5268e77cbfecfcee34780457",
+  "matmul_gemv_subgroup.wgsl#main": "ac84b6dc88fe077dc885d8547e55526bec2f792074dd8746f907ce4a7c342028",
+  "matmul_gemv_subgroup.wgsl#main_multicol": "6631ed8936b6316499e1e1493915dc02a2e137d4f4d2650b62ce63e8805067f1",
+  "matmul_gemv_subgroup.wgsl#main_vec4": "de04e5670494401dd975915e77a603e07144aa1c928c47270afe7a806428cbfd",
   "matmul_gemv.wgsl#main": "dc892efc87edc6d5ddaf191b86c1cc41a603352a332023aa0b1fe55d166673d0",
   "modulate_f16.wgsl#main": "44a98cda1cc7a3575788f865173b9890be792c94e852ac8311b6b8ffbdc1438d",
   "modulate.wgsl#main": "dfe88a35b94752573199c16b3d8aecd4e8e7da57dc88d7b342aa61e0122e71ec",
@@ -182,18 +182,18 @@ export const KERNEL_REF_CONTENT_DIGESTS = Object.freeze({
   "rope.wgsl#rope_ntk_scaled": "818f89865a3d1d6f2d49f671ac882d0fde9709702160a1ae8d9a8ef113afb511",
   "rope.wgsl#rope_qk": "3d773c8b8c400142edc8a4111afb04a2bf75bdb109b2d41cbe5afdb72a959772",
   "rope.wgsl#rope_yarn": "cb00e1cf87fac198dcf0fb0d4e2d5f6f99d2fed6dff0a089a96bb459917851d2",
-  "sample_f16.wgsl#argmax": "30b9f199b49352e5aff91b7aa8016edb423ce33f77481c3a7bc184251856fb27",
-  "sample_f16.wgsl#argmax_reduce": "a3ca27fc50b10c36c1676bdd5dbfe5edc67850cdd5c1af7a1d3ad70f830dd8a7",
-  "sample_f16.wgsl#find_topk_phase1": "24e47e5ced28af802959e350ff0a6eec6b9a26f89fb38e222990eeaffb16bd36",
-  "sample_f16.wgsl#find_topk_phase2": "f40b76d7e1e5606ff6a1e369cf285cb3e802047cfd81377458d4ed0248fcb6a5",
-  "sample_f16.wgsl#sample_single_pass": "f3551288786f0a47f5222e07ab7f54a646e422c2493a62b8074202601eb06ddb",
-  "sample_f16.wgsl#softmax_and_sample": "f1201f01af94c5536463be6439fbe7ed166fcb9494ea081f7d393be3125a7efd",
-  "sample.wgsl#argmax": "bde072faa1ce26e19b6fa1f4ba24c4cd2465909c77a60b175bde1d250113a3b4",
-  "sample.wgsl#argmax_reduce": "b7fa3aea69d9888b9385b380dc70416350d375564b95688592768cd81560b214",
-  "sample.wgsl#find_topk_phase1": "d5fbe00d9ad0c89bd3af72a06f7c964c1b6c2634974ed9a76397fd441422827c",
-  "sample.wgsl#find_topk_phase2": "940b216e605d22096da5aca65950a8030866fc5a39e7fdf484d69a832de1b63a",
-  "sample.wgsl#sample_single_pass": "4412357e84113ee2f1bc0dc8bf89e314c2ab482c89c14ca016ea9949d16a9d0c",
-  "sample.wgsl#softmax_and_sample": "7172c60e76430fbe130e530e3564b569b45eccf193987b32d6f52bd6bbcc9f08",
+  "sample_f16.wgsl#argmax": "7d7188081953a79b6d71bdf783d75df97a78401e2fe62e6d356cc44756a42c41",
+  "sample_f16.wgsl#argmax_reduce": "c0284bc9a50d25e215b21cb8c70e24dae3cb32b578691c2b6df5f3ede68a67c7",
+  "sample_f16.wgsl#find_topk_phase1": "8abd0a978d87adb7cce7337bb1b045a151768724f57802ad060df1dad735cff6",
+  "sample_f16.wgsl#find_topk_phase2": "ea7684cf0cc6014d39ac821edf9c89e140552b5009a72e6e91b00f8816678568",
+  "sample_f16.wgsl#sample_single_pass": "1cd8f0babc5c824b455080d30028109adfe52ce6f79009fbb986fde0d377fcb5",
+  "sample_f16.wgsl#softmax_and_sample": "c1d58cb952b704596d7ab6a2aa32b911a6e869e05b42adac3e4a19d898aa17ae",
+  "sample.wgsl#argmax": "f68b9cfdd3265a5cc52b216e549b629f1f8209e5aaa2f788142fa03db4c2d538",
+  "sample.wgsl#argmax_reduce": "96f8dd75a13db82e1928914e1f40ff1b9e03563eb5f8e3708b230f453b1fc160",
+  "sample.wgsl#find_topk_phase1": "736222d54f805b2791ebb803e9574fca93ab2b25fad0a64245f782499ce2d10c",
+  "sample.wgsl#find_topk_phase2": "a590107f0b7603b4b9624140dea1b436362062f63d64ed6d77e1628578796e77",
+  "sample.wgsl#sample_single_pass": "91c5c30bbc3e034457c1521c1ad576ce798c0868a1fe16e02be5f92706614096",
+  "sample.wgsl#softmax_and_sample": "132d67a1393702c81ee896975447f14f9a6a2589b25125d28401bc8ca47a253d",
   "sana_linear_attention_apply_f16.wgsl#main": "4a7426ce67eccfb70956feeae84275f4d3cc586c50e8442c07eb69993b378ab5",
   "sana_linear_attention_apply.wgsl#main": "5f69e0bc1d9e2df5a61e13bd819313c8f7ff5dfc4b7d78e71d5152dc23b6a86c",
   "sana_linear_attention_summary_f16.wgsl#main": "3abb736ead999485b5dac9c6b534143b464cfd0b5300c5e03c56cec03c8fa48e",
@@ -213,6 +213,8 @@ export const KERNEL_REF_CONTENT_DIGESTS = Object.freeze({
   "softmax.wgsl#main": "45c5876806b442222d7e190e595f55a0079bae82e07d37586996c1a63790bb7a",
   "softmax.wgsl#softmax_online": "6c62601ba2f88f7de9dacf026cc2357168df47d009fd108736655b645217cd0a",
   "softmax.wgsl#softmax_small": "ad75f10e0a3caadd278130504e7d0e4e1b2f0621f8bd390abae5d973e301e47b",
+  "split_qg_f16.wgsl#main": "1d19e9d9900a270a3a1bd407347908f891ba98a1cdcee35ae932257a9a1c72cb",
+  "split_qg.wgsl#main": "64a8aa855c6246675bebdeab0258dc0e649e3986ef2bba4572d2d7dc1af902f4",
   "split_qkv_f16.wgsl#main": "bd1a92fcd9382bdcdf00bcc59248a12489444904a4f10845a381f177b6ad649f",
   "split_qkv.wgsl#main": "bc7c95a47322edc11fec19105efd3774c2adfed151530c849909d03af7503e4c",
   "topk_f16_weights.wgsl#softmax_topk": "863559c28eb46a2b4dc16f21a19aca2424a5d68fc3430b29461bebdd7ec8f625",

package/src/config/kernels/moe/mixtral.paths.json

@@ -0,0 +1,46 @@
+{
+  "id": "mixtral-moe-v1",
+  "description": "Deterministic Mixtral-style MoE kernel routing profile for Q4K/F16 expert weights with standard gate/up/down FFN.",
+  "router": {
+    "topk": [
+      {
+        "match": { "hasF16": true, "hasSubgroups": true, "routerDtype": "f32" },
+        "value": "softmax_topk_f32_subgroup"
+      },
+      {
+        "match": { "hasF16": true, "routerDtype": "f32" },
+        "value": "softmax_topk_f32"
+      },
+      {
+        "match": {},
+        "value": "softmax_topk_f32"
+      }
+    ]
+  },
+  "dequant": {
+    "q4kExpert": [
+      {
+        "match": { "hasF16": true, "hasSubgroups": true, "outputDtype": "f32" },
+        "value": "q4k_expert_dequant_f32_subgroup"
+      },
+      {
+        "match": { "hasF16": true, "outputDtype": "f16" },
+        "value": "q4k_expert_dequant_f16"
+      },
+      {
+        "match": {},
+        "value": "q4k_expert_dequant_f32"
+      }
+    ],
+    "f16Expert": [
+      {
+        "match": { "hasF16": true, "outputDtype": "f16" },
+        "value": "f16_expert_passthrough"
+      },
+      {
+        "match": {},
+        "value": "f16_expert_upcast_f32"
+      }
+    ]
+  }
+}

package/src/config/loader.js

@@ -12,6 +12,7 @@ const transformerPreset = await loadJson('./presets/models/transformer.json', im
 const diffusionPreset = await loadJson('./presets/models/diffusion.json', import.meta.url, 'Failed to load preset');
 const gemma2Preset = await loadJson('./presets/models/gemma2.json', import.meta.url, 'Failed to load preset');
 const gemma3Preset = await loadJson('./presets/models/gemma3.json', import.meta.url, 'Failed to load preset');
+const gemma4Preset = await loadJson('./presets/models/gemma4.json', import.meta.url, 'Failed to load preset');
 const translateGemmaPreset = await loadJson('./presets/models/translategemma.json', import.meta.url, 'Failed to load preset');
 const embeddingGemmaPreset = await loadJson('./presets/models/embeddinggemma.json', import.meta.url, 'Failed to load preset');
 const functiongemmaPreset = await loadJson('./presets/models/functiongemma.json', import.meta.url, 'Failed to load preset');
@@ -24,6 +25,7 @@ const modernbertPreset = await loadJson('./presets/models/modernbert.json', impo
 const lfm2Preset = await loadJson('./presets/models/lfm2.json', import.meta.url, 'Failed to load preset');
 const qwen3Preset = await loadJson('./presets/models/qwen3.json', import.meta.url, 'Failed to load preset');
 const qwen35Preset = await loadJson('./presets/models/qwen3_5.json', import.meta.url, 'Failed to load preset');
+const qwen3VlPreset = await loadJson('./presets/models/qwen3_vl.json', import.meta.url, 'Failed to load preset');
 const kimiK2Preset = await loadJson('./presets/models/kimi-k2.json', import.meta.url, 'Failed to load preset');
 const gptOssPreset = await loadJson('./presets/models/gpt-oss.json', import.meta.url, 'Failed to load preset');
 
@@ -36,6 +38,7 @@ export const PRESET_REGISTRY = {
   transformer: transformerPreset,
   gemma2: gemma2Preset,
   gemma3: gemma3Preset,
+  gemma4: gemma4Preset,
   translategemma: translateGemmaPreset,
   embeddinggemma: embeddingGemmaPreset,
   functiongemma: functiongemmaPreset,
@@ -48,6 +51,7 @@ export const PRESET_REGISTRY = {
   lfm2: lfm2Preset,
   qwen3: qwen3Preset,
   qwen3_5: qwen35Preset,
+  qwen3_vl: qwen3VlPreset,
   kimi_k2: kimiK2Preset,
   gpt_oss: gptOssPreset,
 };
@@ -95,10 +99,12 @@ export const PRESET_DETECTION_ORDER = [
   'diffusion',
   // Model families (check more specific patterns first)
   'gemma2',
+  'gemma4',
   'translategemma',
   'gemma3',
   'llama3',
   'lfm2',
+  'qwen3_vl',
   'qwen3_5',
   'qwen3',
   'kimi_k2',

package/src/config/platforms/loader.js

@@ -9,6 +9,8 @@ const platformCache = new Map();
 
 let platformsBaseUrl = null;
 
+const DEFAULT_PREFER_UNIFIED_MEMORY = false;
+
 const PLATFORM_FILES = [
   'apple-m3',
   'apple-m2',
@@ -131,7 +133,7 @@ export function getMemoryHints() {
 }
 
 export function prefersUnifiedMemory() {
-  return getMemoryHints()?.preferUnifiedMemory ?? false;
+  return getMemoryHints()?.preferUnifiedMemory ?? DEFAULT_PREFER_UNIFIED_MEMORY;
 }
 
 export function getBufferAlignment() {

package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a-nosubgroups.json

@@ -8,19 +8,19 @@
   "decode": {
     "steps": [
       { "op": "input_norm",   "kernel": "rmsnorm.wgsl",         "entry": "main" },
-      { "op": "q_proj",       "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.q_proj" },
-      { "op": "k_proj",       "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.k_proj" },
-      { "op": "v_proj",       "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.v_proj" },
+      { "op": "q_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.q_proj" },
+      { "op": "k_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.k_proj" },
+      { "op": "v_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.v_proj" },
       { "op": "rope_q",       "kernel": "rope.wgsl",            "entry": "main" },
       { "op": "rope_k",       "kernel": "rope.wgsl",            "entry": "main" },
       { "op": "attention",    "kernel": "attention_decode_chunked_f16kv.wgsl", "entry": "main" },
-      { "op": "o_proj",       "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "op": "o_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
       { "op": "attn_residual","kernel": "residual.wgsl",        "entry": "main" },
       { "op": "post_attn_norm","kernel": "rmsnorm.wgsl",        "entry": "main" },
-      { "op": "gate_proj",    "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.mlp.gate_proj" },
-      { "op": "up_proj",      "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.mlp.up_proj" },
+      { "op": "gate_proj",    "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.gate_proj" },
+      { "op": "up_proj",      "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.up_proj" },
       { "op": "activation",   "kernel": "gelu.wgsl",            "entry": "main", "constants": { "HAS_GATE": true } },
-      { "op": "down_proj",    "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
+      { "op": "down_proj",    "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
       { "op": "ffn_residual", "kernel": "residual.wgsl",        "entry": "main" }
     ]
   },
@@ -28,19 +28,19 @@
   "prefill": {
     "steps": [
       { "op": "input_norm",   "kernel": "rmsnorm.wgsl",         "entry": "main" },
-      { "op": "q_proj",       "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.q_proj" },
-      { "op": "k_proj",       "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.k_proj" },
-      { "op": "v_proj",       "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.v_proj" },
+      { "op": "q_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.q_proj" },
+      { "op": "k_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.k_proj" },
+      { "op": "v_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.v_proj" },
       { "op": "rope_q",       "kernel": "rope.wgsl",            "entry": "main" },
       { "op": "rope_k",       "kernel": "rope.wgsl",            "entry": "main" },
       { "op": "attention",    "kernel": "attention_streaming_f16kv.wgsl", "entry": "main" },
-      { "op": "o_proj",       "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "op": "o_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
       { "op": "attn_residual","kernel": "residual.wgsl",        "entry": "main" },
       { "op": "post_attn_norm","kernel": "rmsnorm.wgsl",        "entry": "main" },
-      { "op": "gate_proj",    "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.mlp.gate_proj" },
-      { "op": "up_proj",      "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.mlp.up_proj" },
+      { "op": "gate_proj",    "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.gate_proj" },
+      { "op": "up_proj",      "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.up_proj" },
       { "op": "activation",   "kernel": "gelu.wgsl",            "entry": "main", "constants": { "HAS_GATE": true } },
-      { "op": "down_proj",    "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
+      { "op": "down_proj",    "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
       { "op": "ffn_residual", "kernel": "residual.wgsl",        "entry": "main" }
     ]
   },
@@ -51,8 +51,8 @@
 
   "postLayer": [
     { "op": "final_norm",   "kernel": "rmsnorm.wgsl",          "entry": "main" },
-    { "op": "lm_head",      "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "lm_head" },
-    { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main",  "weights": "lm_head" }
+    { "op": "lm_head",      "kernel": "matmul_f16w_f32a.wgsl",           "entry": "main",  "weights": "lm_head" },
+    { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "lm_head" }
   ],
 
   "sampling": [

package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a-online.json

@@ -28,19 +28,19 @@
   "prefill": {
     "steps": [
       { "op": "input_norm",    "kernel": "rmsnorm.wgsl",                        "entry": "main" },
-      { "op": "q_proj",        "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.self_attn.q_proj" },
-      { "op": "k_proj",        "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.self_attn.k_proj" },
-      { "op": "v_proj",        "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.self_attn.v_proj" },
+      { "op": "q_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.q_proj" },
+      { "op": "k_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.k_proj" },
+      { "op": "v_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.v_proj" },
       { "op": "rope_q",        "kernel": "rope.wgsl",                           "entry": "main" },
       { "op": "rope_k",        "kernel": "rope.wgsl",                           "entry": "main" },
       { "op": "attention",     "kernel": "attention_streaming_f16kv.wgsl",      "entry": "main" },
-      { "op": "o_proj",        "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "op": "o_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
       { "op": "attn_residual", "kernel": "residual.wgsl",                       "entry": "main" },
       { "op": "post_attn_norm","kernel": "rmsnorm.wgsl",                        "entry": "main" },
-      { "op": "gate_proj",     "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.mlp.gate_proj" },
-      { "op": "up_proj",       "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.mlp.up_proj" },
+      { "op": "gate_proj",     "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.gate_proj" },
+      { "op": "up_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.up_proj" },
       { "op": "activation",    "kernel": "gelu.wgsl",                           "entry": "main", "constants": { "HAS_GATE": true } },
-      { "op": "down_proj",     "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
+      { "op": "down_proj",     "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
       { "op": "ffn_residual",  "kernel": "residual.wgsl",                       "entry": "main" }
     ]
   },
@@ -52,7 +52,7 @@
   "postLayer": [
     { "op": "final_norm",      "kernel": "rmsnorm.wgsl",                        "entry": "main" },
     { "op": "lm_head",         "kernel": "matmul_gemv_subgroup.wgsl",           "entry": "main_multicol",  "weights": "lm_head", "constants": { "MULTICOL_COLS_PER_WG": 64, "MULTICOL_THREADS_PER_COL": 4 } },
-    { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a.wgsl",               "entry": "main",  "weights": "lm_head" }
+    { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "lm_head" }
   ],
 
   "sampling": [

package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a-small-attn.json

@@ -0,0 +1,61 @@
+{
+  "id": "gemma3-q4k-dequant-f32a-small-attn",
+  "name": "Gemma 3 Q4K Dequant (F32 activations, small-attn prefill)",
+  "description": "Q4K dequantized to F16 with F32 activations. Same as gemma3-q4k-dequant-f32a-online but uses attention_small_f16kv.wgsl for prefill (diagnostic variant).",
+  "activationDtype": "f32",
+  "kvDtype": "f16",
+
+  "decode": {
+    "steps": [
+      { "op": "input_norm",    "kernel": "rmsnorm.wgsl",                       "entry": "main" },
+      { "op": "q_proj",        "kernel": "matmul_gemv_subgroup.wgsl",           "entry": "main_vec4",  "weights": "layer.{L}.self_attn.q_proj" },
+      { "op": "k_proj",        "kernel": "matmul_gemv_subgroup.wgsl",           "entry": "main_vec4",  "weights": "layer.{L}.self_attn.k_proj" },
+      { "op": "v_proj",        "kernel": "matmul_gemv_subgroup.wgsl",           "entry": "main_vec4",  "weights": "layer.{L}.self_attn.v_proj" },
+      { "op": "rope_q",        "kernel": "rope.wgsl",                           "entry": "main" },
+      { "op": "rope_k",        "kernel": "rope.wgsl",                           "entry": "main" },
+      { "op": "attention",     "kernel": "attention_decode_online_f16kv.wgsl",  "entry": "main" },
+      { "op": "o_proj",        "kernel": "matmul_gemv_subgroup.wgsl",           "entry": "main_vec4",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "op": "attn_residual", "kernel": "residual.wgsl",                       "entry": "main" },
+      { "op": "post_attn_norm","kernel": "rmsnorm.wgsl",                        "entry": "main" },
+      { "op": "gate_proj",     "kernel": "matmul_gemv_subgroup.wgsl",           "entry": "main_vec4",  "weights": "layer.{L}.mlp.gate_proj" },
+      { "op": "up_proj",       "kernel": "matmul_gemv_subgroup.wgsl",           "entry": "main_vec4",  "weights": "layer.{L}.mlp.up_proj" },
+      { "op": "activation",    "kernel": "gelu.wgsl",                           "entry": "main", "constants": { "HAS_GATE": true } },
+      { "op": "down_proj",     "kernel": "matmul_gemv_subgroup.wgsl",           "entry": "main_vec4",  "weights": "layer.{L}.mlp.down_proj" },
+      { "op": "ffn_residual",  "kernel": "residual.wgsl",                       "entry": "main" }
+    ]
+  },
+
+  "prefill": {
+    "steps": [
+      { "op": "input_norm",    "kernel": "rmsnorm.wgsl",                        "entry": "main" },
+      { "op": "q_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.q_proj" },
+      { "op": "k_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.k_proj" },
+      { "op": "v_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.v_proj" },
+      { "op": "rope_q",        "kernel": "rope.wgsl",                           "entry": "main" },
+      { "op": "rope_k",        "kernel": "rope.wgsl",                           "entry": "main" },
+      { "op": "attention",     "kernel": "attention_small_f16kv.wgsl",          "entry": "main" },
+      { "op": "o_proj",        "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.self_attn.o_proj" },
+      { "op": "attn_residual", "kernel": "residual.wgsl",                       "entry": "main" },
+      { "op": "post_attn_norm","kernel": "rmsnorm.wgsl",                        "entry": "main" },
+      { "op": "gate_proj",     "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.gate_proj" },
+      { "op": "up_proj",       "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.up_proj" },
+      { "op": "activation",    "kernel": "gelu.wgsl",                           "entry": "main", "constants": { "HAS_GATE": true } },
+      { "op": "down_proj",     "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "layer.{L}.mlp.down_proj" },
+      { "op": "ffn_residual",  "kernel": "residual.wgsl",                       "entry": "main" }
+    ]
+  },
+
+  "preLayer": [
+    { "op": "embed",           "kernel": "gather_f16.wgsl",                     "entry": "main",  "weights": "embed_tokens" }
+  ],
+
+  "postLayer": [
+    { "op": "final_norm",      "kernel": "rmsnorm.wgsl",                        "entry": "main" },
+    { "op": "lm_head",         "kernel": "matmul_gemv_subgroup.wgsl",           "entry": "main_multicol",  "weights": "lm_head", "constants": { "MULTICOL_COLS_PER_WG": 64, "MULTICOL_THREADS_PER_COL": 4 } },
+    { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a_tiled.wgsl",         "entry": "main",  "weights": "lm_head" }
+  ],
+
+  "sampling": [
+    { "op": "sample",          "kernel": "sample.wgsl",                         "entry": "sample_single_pass" }
+  ]
+}