spec bench: warmup + short prompt + 192-tok decode-dominated (fix prefill confound)

index.html

@@ -144,11 +144,15 @@ const eqArr = (a, b) => a.length === b.length && a.every((x, i) => x === b[i]);
 async function runSpecBench() {
   log.innerHTML = ""; input.disabled = send.disabled = true;
   if (!engine.specAvailable) { st.textContent = "spec-decode unavailable for this model"; bubble("a", "This model can't use the batched-verify head (specAvailable=false)."); return; }
-  const N = 96, rep = m.rep ?? 1.3, rows = [];
+  const N = 192, rep = m.rep ?? 1.3, rows = [];   // long gen so DECODE dominates (prefill was the old confound)
   let prev = { windows: 0, drafted: 0, accepted: 0 };
+  // warm up to boost the GPU clock (cold vs boosted differs ~2.4×) — measure only when warm
+  st.textContent = "warming up (boosting GPU clock)…";
+  globalThis.__spec = false; engine.reset();
+  await engine.generate(engine.tokenize(engine.frameTurn("Write one sentence about the sea.", false)), { maxNew: 48, repPenalty: rep });
   for (const b of SPEC_BENCH) {
     st.textContent = `bench: ${b.tag}…`;
-    const ids = engine.tokenize(frameSystem() + engine.frameTurn(b.text, false));
+    const ids = engine.tokenize(engine.frameTurn(b.text, false));   // NO long system prompt — keep prefill small
     globalThis.__spec = false; engine.reset();
     const t0 = performance.now(); const r0 = await engine.generate(ids.slice(), { maxNew: N, repPenalty: rep }); const w0 = performance.now() - t0;
     globalThis.__spec = true; engine.reset();