webGPU to WASM workaround

src/client.ts

@@ -39,7 +39,7 @@ import type { ModelSourceMode } from "./model-source.js";
 import { applyContextOverride, buildWebLLMAppConfig, TRANSFORMERS_CACHE_KEY } from "./model-source.js";
 import { verifyTransformersProxy, verifyWebLLMProxy } from "./proxy-verify.js";
 import type { HardwareSnapshot } from "./hardware.js";
-import { collectHardwareSnapshot } from "./hardware.js";
+import { collectHardwareSnapshot, isChromiumBased } from "./hardware.js";
 import type { CacheCleanupResult, CachedModelStatus, IndividualModelCleanupResult, ModelCacheTarget, StorageEstimateSnapshot } from "./model-cache.js";
 import {
   deleteOneModelArtifactsFromBrowserStorage,
@@ -261,11 +261,16 @@ export class BrowserAI extends TypedEmitter<BrowserAIEvents> {
     try {
       this.#hardware = await collectHardwareSnapshot();
       let device: MultimodalDevice = "webgpu";
+      const wasmCapable = !!preset.mmRuntime && WASM_CAPABLE_RUNTIMES.has(preset.mmRuntime);
       if (!this.#hardware.webgpuSupported) {
         // Only the runtimes whose loaders honor a wasm device (ASR pipelines, TTS) can fall back;
         // VLM/Gemma/audio-LLM recipes hardcode WebGPU, and text engines require it outright.
-        if (preset.mmRuntime && WASM_CAPABLE_RUNTIMES.has(preset.mmRuntime)) device = "wasm";
+        if (wasmCapable) device = "wasm";
         else throw new WebGPUUnavailableError(this.#hardware.webgpuReason);
+      } else if (wasmCapable && !isChromiumBased()) {
+        // Firefox and Safari now pass the WebGPU probe, but onnxruntime-web's WebGPU backend only
+        // reliably runs on Chromium (shader miscompiles / quantized-session failures elsewhere).
+        device = "wasm";
       }
 
       // Free any loaded models occupying the slots this one needs, waiting out their in-flight runs.

src/hardware.ts

@@ -19,6 +19,21 @@ export type HardwareSnapshot = {
   platform?: string;
 };
 
+/**
+ * Whether this is a Chromium-based browser (Chrome, Edge, Opera, Brave, …), per the UA-Client-Hints
+ * brand list. onnxruntime-web's WebGPU backend is developed and tested against Chromium; Firefox's
+ * and Safari's newer WebGPU implementations still fail on some of its generated WGSL and its
+ * quantized-weight session transforms (errors that only surface at load/inference time), so
+ * wasm-capable runtimes prefer wasm off-Chromium even when the WebGPU probe passes.
+ * navigator.userAgentData is itself Chromium-only, and every Chromium new enough for WebGPU (113+)
+ * ships it, so the brand list is a reliable signal.
+ */
+export function isChromiumBased(): boolean {
+  if (typeof navigator === "undefined") return false;
+  const nav = navigator as Navigator & { userAgentData?: { brands?: Array<{ brand: string }> } };
+  return nav.userAgentData?.brands?.some((entry) => entry.brand === "Chromium") ?? false;
+}
+
 const LIMIT_NAMES = [
   "maxBufferSize",
   "maxStorageBufferBindingSize",

src/multimodal.ts

@@ -214,7 +214,16 @@ export async function loadMultimodal(
         (device === "webgpu"
           ? { encoder_model: "fp32", decoder_model_merged: "q4" }
           : { encoder_model: "fp32", decoder_model_merged: "q8" });
-      const transcriber = await transformers.pipeline("automatic-speech-recognition", preset.id, { device, dtype, progress_callback });
+      const transcriber = await transformers.pipeline("automatic-speech-recognition", preset.id, {
+        device,
+        dtype,
+        progress_callback,
+        // onnxruntime-web 1.26-dev's extended graph optimizations fail at session creation for the
+        // quantized ASR decoders on the wasm EP (qdq_actions.cc: "Missing required scale …
+        // TransposeDQWeightsForMatMulNBits"). Capping at "basic" skips the broken QDQ rewrites —
+        // verified to load and transcribe. The WebGPU path keeps the default level.
+        ...(device === "wasm" ? { session_options: { graphOptimizationLevel: "basic" as const } } : {}),
+      });
       return { kind: "stt", transcriber };
     }
     case "tts-pipeline": {

test/client.test.ts

@@ -4,9 +4,10 @@ import { beforeEach, describe, expect, it, vi } from "vitest";
 /* Module mocks: fake WebGPU probe + fake engines                      */
 /* ------------------------------------------------------------------ */
 
-const hardwareState = { webgpuSupported: true };
+const hardwareState = { webgpuSupported: true, chromiumBased: true };
 
 vi.mock("../src/hardware.js", () => ({
+  isChromiumBased: () => hardwareState.chromiumBased,
   collectHardwareSnapshot: async () => ({
     webgpuSupported: hardwareState.webgpuSupported,
     webgpuReason: hardwareState.webgpuSupported ? "ok" : "no adapter",
@@ -99,6 +100,7 @@ import { UnknownModelError, WebGPUUnavailableError } from "../src/errors.js";
 
 beforeEach(() => {
   hardwareState.webgpuSupported = true;
+  hardwareState.chromiumBased = true;
   created.length = 0;
   transformersCalls.length = 0;
   activeGenerations = 0;
@@ -168,6 +170,17 @@ describe("BrowserAI slot manager", () => {
     expect(call?.options.device).toBe("wasm");
   });
 
+  it("prefers wasm for ASR/TTS presets off-Chromium even when the WebGPU probe passes", async () => {
+    hardwareState.chromiumBased = false;
+    const ai = new BrowserAI();
+    await ai.load("onnx-community/whisper-base");
+    const call = transformersCalls.find((entry) => entry.task === "automatic-speech-recognition");
+    expect(call?.options.device).toBe("wasm");
+    // The wasm EP must cap graph optimization: ORT 1.26-dev's extended QDQ rewrites break the
+    // quantized ASR decoders at session creation ("Missing required scale" / MatMulNBits).
+    expect(call?.options.session_options).toEqual({ graphOptimizationLevel: "basic" });
+  });
+
   it("does NOT offer a wasm fallback for VLM/Gemma/audio-LLM presets (their recipes hardcode WebGPU)", async () => {
     hardwareState.webgpuSupported = false;
     const ai = new BrowserAI();