Upload holo-load2bit.mjs with huggingface_hub

holo-load2bit.mjs

@@ -10,6 +10,7 @@
 // SAME { manifest, fetchTensor } shape — so the engine, KV-cache, and Q's brain loader need no changes.
 import { f16ToF32 } from "./qvac-ingest.mjs";
 import { blake3hex } from "./holo-blake3.mjs";   // BLAKE3 σ-axis κ verify (opt-in via manifest.hash==="blake3") — the GPU-parallel tree hash
+import { gpuBlake3Hex } from "./gpu-blake3.mjs"; // the SAME BLAKE3, run entirely on the GPU (2.74 GB/s) — used when globalThis.__gpuVerify
 
 async function gunzip(u8) { const ds = new DecompressionStream("gzip"); const w = ds.writable.getWriter(); w.write(u8); w.close(); return new Uint8Array(await new Response(ds.readable).arrayBuffer()); }
 const hex = (buf) => [...new Uint8Array(buf)].map((b) => b.toString(16).padStart(2, "0")).join("");
@@ -107,7 +108,13 @@ export async function loadKappaObject(baseUrl, opts = {}) {
     const gz = await cachedBytes(baseUrl + "/b/" + kappa.replace(":", "_") + ".gz");
     // Law L5: re-derive the κ. BLAKE3 (σ-axis, GPU-parallel tree hash) when the manifest declares it — the same
     // hash the on-GPU verifier reproduces; else SHA-256 (SRI axis, default). Untrusted-CDN-safe either way.
-    const got = man.hash === "blake3" ? "blake3:" + blake3hex(gz) : "sha256:" + hex(await crypto.subtle.digest("SHA-256", gz));
+    let got;
+    if (man.hash === "blake3") {
+      if (typeof globalThis !== "undefined" && globalThis.__gpuVerify) {
+        try { got = "blake3:" + await gpuBlake3Hex(gz); }              // re-derive on the GPU (no CPU hash on the hot path)
+        catch (e) { got = "blake3:" + blake3hex(gz); }                 // fall back to CPU if WebGPU verify is unavailable
+      } 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));
     return await gunzip(gz);
   };