warmup uses decode() to precompile batched pipelines (cut first-msg TTFT)

index.html

@@ -115,7 +115,9 @@ async function warmUp() {
   const gpu = engine && engine._gpu;
   if (!gpu || !gpu.sync || !gpu.generate || !gpu.reset) return;
   try {
-    gpu.reset(); await gpu.generate(engine.tokenize("Hello there."), 24, m.rep ?? 1.3); gpu.reset();   // boost clock (throwaway)
+    // throwaway via decode() (not generate()) so it BUILDS the batched-decode pipelines now — else the first real
+    // message pays ~1.7s of one-time shader compilation. Also boosts the GPU clock.
+    gpu.reset(); await (gpu.decode || gpu.generate)(engine.tokenize("Hello there."), 24, m.rep ?? 1.3); gpu.reset();
     let primeIds = engine.tokenize(frameSystem());
     if (m.bos && engine.bosId != null) primeIds = [engine.bosId, ...primeIds];
     await gpu.sync(primeIds, true);                        // prefill the system prompt into KV; leaves cache == primeIds