Upload fork.js with huggingface_hub

fork.js

@@ -539,18 +539,26 @@ async function search(append) {
     search(true);
     return;
   }
-  // best-sort promises fitting models first, but the server can no longer
-  // pre-filter (it only speaks f16): order client-side, full offloads
-  // first, then partials, each by upstream score. count what is shown —
-  // the server total is unfiltered while rescaling.
+  // the server can no longer pre-filter fits (it only speaks q8_0/f16),
+  // so best-sort keeps server score order and the client drops whatever
+  // has no pick. count what is shown — the server total is unfiltered.
+  // the vram-only and fits-only checkboxes are enforced here too, since
+  // their server-side equivalents are forced off while rescaling.
+  if (bestSort || $("#fits-only").checked) {
+    for (const m of seenModels) {
+      if (!m._ui || !m._ui.c.isConnected) continue;
+      if (!m._pick) { m._ui.c.remove(); nofitDropped++; }
+    }
+  }
+  if ($("#vram-only").checked && r.kind === "gpu") {
+    for (const m of seenModels) {
+      if (!m._ui || !m._ui.c.isConnected) continue;
+      if (!m._pick || !m._pick.fit || m._pick.fit.tier !== "gpu") {
+        m._ui.c.remove(); nofitDropped++;
+      }
+    }
+  }
   if (bestSort) {
-    // whole-resident (gpu/cpu/unified/ram) before experts-in-RAM and
-    // partials: gpu_moe counts as full in the ranker but is not whole.
-    const fullFirst = (m) => (m._pick && m._pick.full &&
-      m._pick.fit && m._pick.fit.tier !== "gpu_moe" ? 0 : 1);
-    seenModels.filter((m) => m._ui && m._ui.c.isConnected)
-      .sort((a, b) => fullFirst(a) - fullFirst(b) || b.score - a.score)
-      .forEach((m) => box.append(m._ui.c));
     $("#count").textContent =
       `${box.childElementCount.toLocaleString()} models that fit your hardware`;
   }