Upload holo-load2bit.mjs with huggingface_hub

holo-load2bit.mjs

@@ -110,12 +110,12 @@ export async function loadKappaObject(baseUrl, opts = {}) {
     // hash the on-GPU verifier reproduces; else SHA-256 (SRI axis, default). Untrusted-CDN-safe either way.
     let got;
     if (man.hash === "blake3") {
-      // OPT-IN (?verify=gpu → globalThis.__gpuVerify): re-derive on the GPU. Default = CPU BLAKE3 (the proven,
-      // working path). GPU verify wired here fails during engine construction (2-device conflict — see note);
-      // needs a single-device refactor before it can be default. On any GPU error, fall back to CPU.
-      if (typeof globalThis !== "undefined" && globalThis.__gpuVerify && gpuBlake3Available()) {
-        try { const h = await gpuBlake3Hex(gz); got = "blake3:" + h; }
-        catch (e) { got = "blake3:" + blake3hex(gz); }
+      // DEFAULT: re-derive on the SHARED GPU device (2.74 GB/s). SAFE fallback: on any GPU error OR if the GPU
+      // digest disagrees with the pin, re-derive on the CPU — so a GPU quirk degrades to CPU speed, never fails
+      // the load. Only a CPU mismatch (real corruption) throws. Set globalThis.__gpuVerify=false to force CPU.
+      if (gpuBlake3Available() && (typeof globalThis === "undefined" || globalThis.__gpuVerify !== false)) {
+        try { got = "blake3:" + await gpuBlake3Hex(gz); } catch (e) { got = "blake3:" + blake3hex(gz); }
+        if (got !== kappa) got = "blake3:" + blake3hex(gz);   // GPU disagreed → trust the CPU verifier
       } else got = "blake3:" + blake3hex(gz);
     } else got = "sha256:" + hex(await crypto.subtle.digest("SHA-256", gz));
     if (got !== kappa) throw new Error("κ MISMATCH " + kappa.slice(0, 24));