feat: enable acceleration via vulkan

.github/workflows/release.yml

@@ -5,6 +5,7 @@ on:
     branches:
       - feat.windows.support
       - feat/whisper.migration
+      - feat/vulkan-support
     tags:
       - 'v*'
   workflow_dispatch:
@@ -71,12 +72,12 @@ jobs:
           node-version: '24.1.0'
           cache: 'pnpm'
 
-      - name: Install CUDA Toolkit
+      - name: Install Vulkan SDK
         if: matrix.os == 'windows'
-        uses: Jimver/cuda-toolkit@v0.2.15
+        uses: humbletim/install-vulkan-sdk@v1.2
         with:
-          cuda: '12.4.1'
-          method: 'network'
+          version: 1.3.290.0
+          cache: true
 
       - name: Log Node.js architecture and platform
         run: |
@@ -98,11 +99,11 @@ jobs:
           GGML_NATIVE: OFF # CI mac runners lack i8mm support; keep CPU features conservative here
         run: pnpm --filter @amical/whisper-wrapper build:native
 
-      - name: Build whisper native binaries (cuda)
+      - name: Build whisper native binaries (vulkan)
         if: matrix.os == 'windows'
         env:
           GGML_NATIVE: OFF
-        run: pnpm --filter @amical/whisper-wrapper build:native:cuda
+        run: pnpm --filter @amical/whisper-wrapper build:native:vulkan
 
       - name: Download Node.js binaries
         working-directory: apps/desktop

packages/whisper-wrapper/WHISPER_CPP_VERSION

@@ -1 +1 @@
-v1.7.6
+v1.8.2

packages/whisper-wrapper/bin/build-addon.js

@@ -185,6 +185,9 @@ function variantFromName(name, platform, arch) {
   if (name.includes("-cuda")) {
     envOverrides.GGML_CUDA = "1";
   }
+  if (name.includes("-vulkan")) {
+    envOverrides.GGML_VULKAN = "1";
+  }
   if (name.startsWith("darwin-")) {
     envOverrides.GGML_USE_ACCELERATE = envOverrides.GGML_USE_ACCELERATE || "1";
   }
@@ -262,6 +265,12 @@ for (const variant of variants) {
   };
 
   propagateCMakeBool("GGML_NATIVE");
+  propagateCMakeBool("GGML_VULKAN");
+  propagateCMakeBool("GGML_METAL");
+  propagateCMakeBool("GGML_CUDA");
+  propagateCMakeBool("GGML_OPENBLAS");
+  propagateCMakeBool("GGML_BLAS");
+  propagateCMakeBool("GGML_USE_ACCELERATE");
 
   run(cmakeParts.join(" "), {
     cwd: addonDir,

packages/whisper-wrapper/package.json

@@ -19,7 +19,8 @@
     "build": "tsc -p tsconfig.json",
     "postinstall": "node ./bin/build-addon.js",
     "build:native": "node ./scripts/build-native.js",
-    "build:native:cuda": "node ./scripts/build-native.js --cuda"
+    "build:native:cuda": "node ./scripts/build-native.js --cuda",
+    "build:native:vulkan": "node ./scripts/build-native.js --vulkan"
   },
   "dependencies": {
     "cmake-js": "^7.3.1",

packages/whisper-wrapper/scripts/build-native.js

@@ -12,7 +12,9 @@ function build(targets) {
   });
 }
 
-if (process.argv.includes("--cuda")) {
+if (process.argv.includes("--vulkan")) {
+  build(["win32-x64-vulkan", "win32-x64"]);
+} else if (process.argv.includes("--cuda")) {
   build(["win32-x64-cuda", "win32-x64"]);
 } else {
   build([]);

packages/whisper-wrapper/src/loader.ts

@@ -1,7 +1,7 @@
 import path from "node:path";
 import fs from "node:fs";
 
-const GPU_FIRST_CANDIDATES = ["metal", "openblas", "cuda"] as const;
+const GPU_FIRST_CANDIDATES = ["metal", "openblas", "cuda", "vulkan"] as const;
 
 function candidateDirs(platform: string, arch: string): string[] {
   return [
@@ -66,13 +66,15 @@ export function loadBinding(): any {
       // Store the loaded binding info
       const bindingType = dir.includes("-cuda")
         ? "cuda"
-        : dir.includes("-metal")
-          ? "metal"
-          : dir.includes("-openblas")
-            ? "openblas"
-            : dir === "cpu-fallback"
-              ? "cpu-fallback"
-              : "cpu";
+        : dir.includes("-vulkan")
+          ? "vulkan"
+          : dir.includes("-metal")
+            ? "metal"
+            : dir.includes("-openblas")
+              ? "openblas"
+              : dir === "cpu-fallback"
+                ? "cpu-fallback"
+                : "cpu";
       loadedBindingInfo = {
         path: candidate,
         type: bindingType,

packages/whisper-wrapper/whisper.cpp

@@ -1 +1 @@
-Subproject commit a8d002cfd879315632a579e73f0148d06959de36
+Subproject commit 4979e04f5dcaccb36057e059bbaed8a2f5288315