llama serve rename, 3-bit default floor, blank dominated warnings, band-worst heuristics, hide non-fits on best-sort

fork.js

@@ -6,7 +6,7 @@
 const API = "https://codelion-local-model-explorer.hf.space";
 const $ = (s, r = document) => r.querySelector(s);
 const $$ = (s, r = document) => [...r.querySelectorAll(s)];
-let META = null, SCORES = null, runId = 0, offset = 0, total = 0;
+let META = null, SCORES = null, runId = 0, offset = 0, total = 0, nofitDropped = 0;
 
 function el(tag, cls, text) {
   const n = document.createElement(tag);
@@ -59,7 +59,7 @@ function query(append) {
     sort: $("#sort").value, family: $("#family").value || null, size: $("#size").value || null,
     search: $("#search").value.trim() || null,
     moe_only: $("#moe-only").checked, vision_only: $("#vision-only").checked,
-    fits_only: $("#fits-only").checked,
+    fits_only: $("#fits-only").checked || $("#sort").value === "best",
     vram_only: r.kind === "gpu" && $("#vram-only").checked, mmproj: $("#mmproj").value,
     offset: append ? offset : 0, limit: 24,
   };
@@ -147,7 +147,8 @@ function renderDetail(m, body) {
   if (sel && sel.commands) {
     const row = el("div", "row");
     const copy = el("button", "btn ghost", "Copy command");
-    const pre = el("pre", "cmd", sel.commands["llama.cpp"]);
+    // upstream still emits the old `llama-server` binary name; current llama.cpp renamed it.
+    const pre = el("pre", "cmd", (sel.commands["llama.cpp"] || "").replace(/llama-server/g, "llama serve"));
     copy.addEventListener("click", async () => {
       try { await navigator.clipboard.writeText(pre.textContent); copy.textContent = "Copied"; }
       catch { copy.textContent = "Select and copy"; }
@@ -360,6 +361,10 @@ async function search(append) {
   $("#count").textContent = `${total.toLocaleString()} models`;
   $("#more").hidden = offset >= total;
   const r = rig();
+  // best-sort promises fitting models, so non-fits are filtered server-side
+  // (fits_only above) and dropped client-side (headroom residuals) below.
+  const bestSort = q.sort === "best";
+  if (!append) nofitDropped = 0;
 
   for (const m of res.results) {
     const ui = card(m);
@@ -395,6 +400,9 @@ async function search(append) {
           if (hr) ui.status.append(el("div", "muted", hr));
           const railEl = rail(pick.fit, r);
           if (railEl) ui.status.append(railEl);
+        } else if (bestSort) {
+          ui.c.remove();
+          nofitDropped++;
         } else {
           ui.status.textContent = warnings[0] || "Doesn't fit this hardware.";
         }
@@ -405,6 +413,8 @@ async function search(append) {
     }
   }
   await Promise.all(Array.from({ length: 4 }, worker));
+  if (alive() && bestSort) $("#count").textContent =
+    `${(total - nofitDropped).toLocaleString()} models that fit your hardware`;
 }
 
 let deb;

index.html

@@ -103,7 +103,7 @@
         <option value="q4_0">4-bit (q4_0) — bench: small KL cost</option>
       </select></div>
       <div class="field"><label for="mmproj">Vision projector (vision models)</label><select id="mmproj">
-        <option value="gpu">On the GPU, llama-server default</option>
+        <option value="gpu">On the GPU, llama serve default</option>
         <option value="ram">In system RAM (--no-mmproj-offload)</option>
         <option value="off">Don't load it, text only (--no-mmproj)</option>
       </select></div>
@@ -122,7 +122,7 @@
       <div class="field"><label for="min-bits">Lowest quant to consider</label><select id="min-bits">
         <option value="0">Any, down to 1-bit</option>
         <option value="2">2-bit and up</option>
-        <option value="3">3-bit and up</option>
+        <option value="3" selected>3-bit and up</option>
         <option value="4">4-bit and up (Q4, IQ4, MXFP4)</option>
         <option value="5">5-bit and up</option>
         <option value="6">6-bit and up</option>

rankQuants.js

@@ -7,8 +7,8 @@
  * recommend/model handlers; fit estimation stays untouched.
  *
  * Single mode: best measured quality that sits fully on the hardware
- * wins, with warnings when nothing fits fully / only sub-4-bit fits /
- * the pick is a legacy format. no band caps: if Q8_0 fits, it wins.
+ * wins, with warnings when nothing fits fully / only sub-4-bit fits.
+ * no band caps: if Q8_0 fits, it wins.
  *
  * Fit awareness: full offloads (tier gpu/gpu_moe/cpu/unified/ram) always
  * outrank partial ones. A partial "quality" pick (e.g. 12 of 64 layers on
@@ -52,11 +52,14 @@ function scoreMap(scoresJson) {
 // measurement; ties break measured-first in the sort below.
 function classifyUnknown(quant) {
   const q = quant.toUpperCase();
-  if (/Q4_0|Q4_1|Q3_K_S|Q3_K_M|Q5_K_S|Q4_K_S|Q4_K_L|Q5_0|Q3_0/.test(q)) return ["avoid", 0.10];
+  // heuristic kl must equal the WORST measured kl in its band (avoid
+  // 0.5345 = UD-Q2_K_XL, emergency 1.0 = UD-IQ2_XXS; see quant_scores.json)
+  // so a guess never outranks a measurement; ties break measured-first.
+  if (/Q4_0|Q4_1|Q3_K_S|Q3_K_M|Q5_K_S|Q4_K_S|Q4_K_L|Q5_0|Q3_0/.test(q)) return ["avoid", 0.5345];
   const bit = q.match(/(\d(\.\d)?)\s*BIT/);
-  if (bit) return parseFloat(bit[1]) < 4 ? ["emergency", 0.50] : ["tight", 0.1166];
-  if (/IQ1|1BIT|_I1|TERNARY|Q1(?!\d)/.test(q)) return ["emergency", 0.60];
-  if (/Q2|IQ2/.test(q)) return ["emergency", 0.40];
+  if (bit) return parseFloat(bit[1]) < 4 ? ["emergency", 1.0] : ["tight", 0.1166];
+  if (/IQ1|1BIT|_I1|TERNARY|Q1(?!\d)/.test(q)) return ["emergency", 1.0];
+  if (/Q2|IQ2/.test(q)) return ["emergency", 1.0];
   if (/IQ3|Q3/.test(q)) return ["tight", 0.1166];
   if (/Q4|IQ4|MXFP4/.test(q)) return ["default", 0.0541];
   if (/Q5|Q6/.test(q)) return ["step", 0.0118];
@@ -164,9 +167,6 @@ function rankOptions(options, family, scoresJson) {
 
   // best measured quality that fits. no band caps.
   let pick = byKlThenPub(cands, family)[0];
-  if (pick.band === "avoid") {
-    warnings.push(`legacy pick: ${pick.quant} is an outdated format — nothing better in this repo`);
-  }
 
   const ranked = byKlThenPub(cands, family);
   const twins = cands.filter((o) => o.quant.toUpperCase() === pick.quant.toUpperCase());