Upload holo-load2bit.mjs with huggingface_hub

holo-load2bit.mjs

@@ -43,8 +43,12 @@ async function cachedBytes(url) {
 // FAST FIRST LOAD: warm the whole block cache with bounded concurrency, so the engine's sequential per-tensor
 // reads hit the cache instead of paying one HF round-trip at a time (~1.5 blocks/s → tens of blocks/s over HTTP/2).
 // Fire-and-forget; the engine's getBlock shares any in-flight fetch (no double-download). Cross-origin-CDN safe.
-function prefetchBlocks(baseUrl, kappas, conc = 12) {
+function prefetchBlocks(baseUrl, kappas, conc) {
   try {
+    // HF serves over HTTP/2: many small blocks parallelize well, so fan out wide to saturate bandwidth (the one
+    // big embed block holds a single lane, the rest fill the others). Tunable via globalThis.__prefetchConc for
+    // constrained links/mobile. This is the whole first-load story for a resident κ-object — bandwidth, not RTT.
+    conc = conc || (typeof globalThis !== "undefined" && globalThis.__prefetchConc) || 24;
     const urls = [...new Set(kappas.filter(Boolean))].map((k) => baseUrl + "/b/" + String(k).replace(":", "_") + ".gz");
     let i = 0;
     const worker = async () => { while (i < urls.length) { const u = urls[i++]; try { await cachedBytes(u); } catch (e) {} } };