keep-warm pulse: hold GPU boost clock while typing so casual chat decodes at boosted rate

index.html

@@ -141,6 +141,41 @@ send.onclick = onSend;
 input.onkeydown = (e) => { if (e.key === "Enter" && !e.shiftKey) { e.preventDefault(); onSend(); } };
 input.oninput = () => { input.style.height = "auto"; input.style.height = Math.min(140, input.scrollHeight) + "px"; };
 
+// KEEP-WARM: casual chat lets the GPU cool between messages, so each reply decodes at a medium clock (~40ms/tok)
+// instead of boosted (~20ms). While the user is typing/focused, run a tiny THROWAWAY GPU compute (scratch buffer,
+// never touches the model KV) to hold the clock up → the next message decodes boosted. Pauses during generation
+// and when the tab is hidden. This is the idle-cooling fix for sustained high tok/s in normal use.
+let _warmDev = null, _warmPipe = null, _warmBg = null, _warmBusy = false, _warmTimer = null;
+function ensureWarmKit() {
+  if (_warmPipe) return true;
+  const dev = engine && engine._gpu && engine._gpu._dev && engine._gpu._dev();
+  if (!dev) return false;
+  _warmDev = dev;
+  const WGSL = `@group(0) @binding(0) var<storage,read_write> b: array<vec4<f32>>;
+@compute @workgroup_size(64) fn main(@builtin(global_invocation_id) g:vec3<u32>){
+  let i = g.x % 65536u; var a = b[i]; for(var k=0u;k<1024u;k++){ a = a*1.0000001 + vec4<f32>(0.5); } b[i] = a; }`;
+  try {
+    _warmPipe = dev.createComputePipeline({ layout: "auto", compute: { module: dev.createShaderModule({ code: WGSL }), entryPoint: "main" } });
+    const buf = dev.createBuffer({ size: 65536 * 16, usage: GPUBufferUsage.STORAGE });
+    _warmBg = dev.createBindGroup({ layout: _warmPipe.getBindGroupLayout(0), entries: [{ binding: 0, resource: { buffer: buf } }] });
+  } catch (e) { console.warn("warmkit", e); return false; }
+  return true;
+}
+async function warmPulse(ms = 220) {
+  if (busy || _warmBusy || document.hidden || !armed) return;   // never compete with real inference
+  if (!ensureWarmKit()) return;
+  _warmBusy = true;
+  try {
+    const end = performance.now() + ms;
+    while (performance.now() < end && !busy && !document.hidden) {
+      const e = _warmDev.createCommandEncoder(); const p = e.beginComputePass(); p.setPipeline(_warmPipe); p.setBindGroup(0, _warmBg); p.dispatchWorkgroups(4096); p.end(); _warmDev.queue.submit([e.finish()]);
+      await _warmDev.queue.onSubmittedWorkDone();
+    }
+  } catch (e) { /* ignore — best effort */ } finally { _warmBusy = false; }
+}
+input.addEventListener("focus", () => warmPulse(300));
+input.addEventListener("keydown", () => { clearTimeout(_warmTimer); _warmTimer = setTimeout(() => warmPulse(220), 50); });
+
 // ── SPEC-DECODE A/B HARNESS (?bench=spec) ── measures baseline greedy vs speculative on the operator's real
 // GPU across echo-heavy and free-form prompts: byte-identical check (G1), mean accepted tokens/verify (G2),
 // and decode tok/s for both. One load, one table — the honest verdict on whether spec-decode earns its place.