Upload holo-gpu-device.mjs with huggingface_hub

holo-gpu-device.mjs

@@ -0,0 +1,27 @@
+// holo-gpu-device.mjs — ONE WebGPU device for the whole app (the engine, BLAKE3 verify, Bao streaming).
+// Memoized: exactly one requestDevice() per page. Created with the ADAPTER'S MAX limits so a 145 MB block
+// binds as a storage buffer (the default cap is 128 MB — a separate default-limits device silently produced a
+// wrong hash on the embed → κ mismatch → the two-device clash). timestamp-query is requested when available.
+// Sharing ONE device is what lets fetch→verify→decode coexist without contending.
+let _p = null;
+export function getDevice() {
+  if (_p) return _p;
+  _p = (async () => {
+    if (typeof navigator === "undefined" || !navigator.gpu) throw new Error("no WebGPU");
+    const adapter = await navigator.gpu.requestAdapter();
+    const L = adapter.limits;
+    const canTs = adapter.features.has("timestamp-query");
+    const dev = await adapter.requestDevice({
+      requiredFeatures: canTs ? ["timestamp-query"] : [],
+      requiredLimits: {
+        maxStorageBufferBindingSize: L.maxStorageBufferBindingSize,
+        maxBufferSize: L.maxBufferSize,
+        maxComputeWorkgroupsPerDimension: L.maxComputeWorkgroupsPerDimension,
+      },
+    });
+    try { if (typeof globalThis !== "undefined") globalThis.__gpuDeviceCount = (globalThis.__gpuDeviceCount || 0) + 1; } catch (e) {}
+    return { dev, adapter, canTs };
+  })().catch((e) => { _p = null; throw e; });
+  return _p;
+}
+export default { getDevice };